Maternal and infant product dynamic recommendation method and system based on user behaviors

Abstract

The invention discloses a maternal and infant product dynamic recommendation method and system based on user behaviors, and the method comprises the following steps: collecting the multi-dimensional behavior data of a user in real time at a client, carrying out the cleaning and feature extraction, and carrying out the structural storage; dynamically establishing and correcting the life cycle state of the user according to the expected date of delivery or the birth date of the infant input by the user; weighting the behavior data according to a time interval, a type and a scene label, and calculating a behavior feature weight; generating a behavior vector based on the weight, and screening similar users in combination with a stage synchronism regulation factor; and integrating the historical purchase behaviors of the similar users and the behaviors of the target users, calculating recommendation scores of the products, and sequencing to generate a personalized recommendation list. According to the method, dynamic modeling can be carried out based on the multi-dimensional behaviors and the life cycle of the user, stage accurate recommendation and real-time personalized pushing of mother and baby products are achieved, and therefore recommendation relevance and user experience are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce recommendation systems, and in particular to a method and system for dynamically recommending maternal and infant products based on user behavior. Background Art

[0002] Existing maternal and infant e-commerce platforms and vertical communities generally use recommendation algorithms such as collaborative filtering, which primarily calculate product similarity based on users' historical purchase data. However, this type of static recommendation strategy fails to consider the short life cycle and rapidly changing needs of maternal and infant users. Users' needs frequently change as their babies grow, such as transitioning from stage 1 to stage 2 formula. However, traditional algorithms, which do not incorporate a time decay factor, often experience a significant lag in recommendation results. For example, field tests have found that some users' needs have changed over time, but the recommendation system continues to push expired products for a long time, with a lag of more than 30 days, affecting user experience and product conversion efficiency.

[0003] Existing recommendation technologies have obvious limitations in user behavior modeling. Most systems only collect and utilize users' historical purchasing behaviors, and lack the mining and analysis of multi-dimensional behavioral characteristics such as browsing, adding to cart, staying, and searching. In the maternal and infant scenario, user behavior has stronger stage and scenario characteristics. A single purchase behavior is difficult to fully reflect the real needs, resulting in a decrease in the relevance and accuracy of the recommendation system. At the same time, the existing recommendation system lacks a dynamic update mechanism for real-time user behavior, and cannot adjust the recommendation strategy in time according to the user's sudden search or temporary consultation operations, thus missing the opportunity for scenario-based recommendation. In practice, the system's response to key behaviors is delayed by more than 5 minutes, which cannot meet the needs of maternal and infant users for instant recommendations.

[0004] Therefore, how to provide a dynamic recommendation method and system for maternal and infant products based on user behavior is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0005] One purpose of the present invention is to propose a dynamic recommendation method and system for maternal and infant products based on user behavior. The present invention significantly improves the maternal and infant product recommendation system's ability to perceive and respond to users' stage-by-stage needs and real-time behavioral changes through multi-dimensional behavioral data collection, dynamic user life cycle modeling, and a stage-by-stage weighting mechanism. The system can not only comprehensively integrate users' basic behavior, implicit behavior, and scenario behavior data, but also dynamically adjust the recommendation strategy based on the user's current pregnancy or infant growth stage, thereby achieving stage accuracy and scenario adaptability in product recommendations. The recommendation model integrates multi-source signals such as collaborative filtering scores, historical behavior scores, and stage matching scores, and uses a weighted hybrid algorithm for comprehensive sorting, effectively improving the relevance and real-time nature of the recommendation results. Practical applications have shown that the present invention can significantly reduce the recommendation of outdated or irrelevant products, improve user satisfaction and recommendation conversion rates, and has good commercial promotion and industrial application value.

[0006] The method for dynamically recommending maternal and infant products based on user behavior according to an embodiment of the present invention includes the following steps: Collect users' multi-dimensional behavior data in real time on the client side, perform data cleaning and feature extraction on the multi-dimensional behavior data, and store it in a structured data format to generate structured multi-dimensional behavior data; Dynamically establish the user lifecycle status based on the user's input due date or baby's birth date and multi-dimensional behavioral data, and perform data correction through a combination of finite state machine and Bayesian inference mechanism; For structured multidimensional behavior data, we use the time interval, behavior type, and current scene label to weight the behavior feature data using the time series decay factor, behavior type weight, and scene relevance parameter to calculate the behavior feature weight. Based on the weights of behavioral features in the most recent period, a behavioral vector is generated according to product category. Combined with the stage synchronization adjustment factor, weighted cosine similarity is used to calculate the similarity between the user and the reference user group to screen similar user sets; Based on the historical purchasing behavior of similar users and the user's own behavior, the collaborative filtering score, historical behavior score and stage matching score are calculated respectively. A weighted hybrid model is used to weight the collaborative filtering score, historical behavior score and stage matching score to generate a recommendation score for each target product. The products are then sorted by the recommendation score, abnormal products are excluded, and the final recommendation list is output.

[0007] Furthermore, the multidimensional behavioral data includes basic behaviors, implicit behaviors and scenario behaviors. Basic behaviors include browsing products, adding to cart, collecting, purchasing, and commenting. Implicit behaviors include page dwell time, scrolling depth, number of image magnifications, and click heat value. Scenario behaviors include user search keywords, consulting content, and browsing paths. The data cleaning includes filtering abnormal data, including browsing with a dwell time of <1 second, browsing with a scrolling depth of <10%, and repeated clicks of the same IP >10 times within one minute. The feature extraction uses an LSTM model to extract keywords from search terms and map them to a preset scenario tag library.

[0008] Furthermore, the data correction step includes: When a user first registers, they enter the due date or the baby's birth date to generate the initial life cycle stage; Based on the passage of time, the current gestational age or baby's age is calculated and updated every 24 hours; Verify based on user-specific behavior. When it is detected that a user purchases a product corresponding to a specific stage, the infant stage is corrected to the product's applicable range. When the life cycle stage calculated based on time is inconsistent with the user's actual behavior data, a pop-up prompt is triggered to prompt the user to confirm whether the baby's current stage needs to be corrected. After the user confirms, the life cycle stage is updated through the state machine mechanism.

[0009] Furthermore, the step of calculating the behavioral feature weight includes: The user's various behavioral data are weighted according to the time interval between behaviors using a time series decay function to obtain a time series decay factor. According to the current life cycle stage of the user, the product category is assigned a scenario relevance parameter to obtain a scenario weight factor, which is set by the key nodes of the life cycle; The behavior type weight is obtained based on the statistics of the user's historical behavior data, specifically including: Collect massive amounts of user behavior data covering a certain period, including browsing, adding to cart, purchasing, favorites, and comments; For each behavior type, calculate the probability of a user ultimately converting to a purchase after completing the behavior. Using the conversion probability of browsing behavior as the benchmark value, calculate the conversion contribution coefficients of other behavior types relative to browsing behavior; Use normalization to adjust the contribution of all behavior types to the range of 0-1, and conduct A / B testing in actual business. The behavior type weight, temporal decay factor, and scenario weight factor are multiplied together to obtain the final behavior feature weight.

[0010] Furthermore, the calculation process of the timing attenuation factor specifically includes: Determine the time interval between a user's behavior and the current one; Query the weight window period corresponding to the behavior type; If the time interval between the occurrence of the behavior and the current weight window period is less than the current weight window period, the behavior's timing decay factor is set to 1. During this window period, the behavior's timing decay factor will not be decayed and the maximum weight will always be maintained. If the time interval between the occurrence of the behavior and the weight window period has exceeded, the temporal attenuation factor of the behavior is calculated as follows: Calculate the duration of time the behavior occurs outside the window period; According to the preset decay rate parameters of the behavior type, an exponential decay method is adopted. The specific calculation is to use the decay rate as the base, take the duration exceeding the window period as the exponent, and obtain a value less than 1 as the temporal decay factor of the behavior.

[0011] Furthermore, the step of screening the similar user set includes: All behavioral data of target users in the most recent preset period is classified, counted, and aggregated by product category to form a behavioral feature vector based on each core maternity and baby product category. Generate behavioral feature vectors for all candidate users in the same way and calculate weighted cosine similarity with the target user's vector; According to the weighted cosine similarity, the top N users with similarity are selected from all candidate users as the similar user set of the target user.

[0012] Furthermore, the calculation process of the weighted cosine similarity specifically includes: In each category dimension of the behavioral feature vector, assign stage-related weights to each product category based on the current life cycle stage of the target user and candidate user; Before calculating the cosine similarity, the behavior vectors of the target user and the candidate user are weighted and normalized to obtain the stage-weighted behavior vectors; Introducing a phase synchronization adjustment factor, only when the target user and the candidate user's category activity period overlap, will the set similarity score be given; In the final similarity score, the weighted cosine value is multiplied by the phase synchronization adjustment factor to obtain the similarity calculation result.

[0013] Furthermore, the step of outputting the recommendation list specifically includes: For each product to be recommended, the total number of users in the similar user set who have purchased the product to be recommended is counted. The number of users who actually purchased the product to be recommended is divided by the total number of similar user sets to obtain the collaborative filtering score of the product to be recommended. For each product to be recommended, summarize all the target user's recent behaviors related to the product to be recommended, multiply them by their respective behavioral feature weights, add up the weighted results and normalize them to obtain the historical behavior score of the product to be recommended; Obtain the time interval of the current user's life cycle stage and the product's applicable life cycle interval, calculate the overlap duration of the two time intervals, divide the overlap duration by the total duration of the product's applicable interval to obtain the stage overlap ratio, and multiply the stage overlap ratio by the preset stage matching correction coefficient to obtain the final stage matching score; For each product to be recommended, the collaborative filtering score, historical behavior score, and stage matching score are weighted and summed according to the preset weights to serve as the final comprehensive recommendation score for the product to be recommended. The products are then sorted from high to low based on the comprehensive recommendation scores to generate a recommendation list. The products in the recommendation list are checked one by one, and products that have been purchased by the user are eliminated. At the same time, products with prices higher than twice the average price of the user's historical purchases are eliminated. Jaccard similarity is further used to compare the user's historical purchase price set with the set of product prices to be recommended. When the price deviation is greater than the set threshold, the product is not recommended, and the final recommendation list is generated.

[0014] The dynamic recommendation system for maternal and infant products based on user behavior according to an embodiment of the present invention includes the following modules: The multi-dimensional behavior data collection module is used to collect multi-dimensional behavior data of users, including basic behavior, implicit behavior and scenario behavior data, from the client in real time, and perform data cleaning and feature extraction on the collected multi-dimensional behavior data, and store it in a structured data format; A user lifecycle modeling module is used to establish a dynamic state machine of the user lifecycle based on the due date or baby birth date information input by the user and the behavioral data uploaded by the multi-dimensional behavioral data acquisition module, and to correct the lifecycle state and switch stages using finite state automata and Bayesian inference rules; The behavior feature weight calculation module is used to calculate the behavior feature weight by weighting the collected multi-dimensional behavior data according to the time interval of the behavior, the behavior type and the current scene label using the time series attenuation factor, the behavior type weight and the scene relevance parameter respectively; The dynamic similarity calculation module is used to aggregate the weights of users' behavioral features by product category based on a set time period to construct a behavior vector. This vector is then combined with a stage synchronization adjustment factor to analyze the similarity of the target user's behavior patterns with other users in the same or adjacent lifecycle stages using weighted cosine similarity, thereby screening out a set of similar users. The recommendation result generation module is used to integrate three types of scoring results: collaborative filtering score, user historical behavior score and stage matching score, and weighted summary according to the preset hybrid weight model to complete the comprehensive recommendation score sorting of the target products, sort them by recommendation score, exclude abnormal products, and output the final recommendation list.

[0015] The beneficial effects of the present invention are: This invention introduces a multi-dimensional behavioral data collection and real-time feature modeling mechanism, effectively overcoming the lag and single-dimensionality issues of traditional collaborative filtering recommendations in maternal and infant scenarios. The system not only collects basic user behavioral data but also integrates implicit and scenario-based behaviors, achieving a comprehensive and accurate perception of user interest dynamics and stage needs. Combining the time series decay factor, behavior type weight, and scenario relevance factor, it dynamically weights the freshness of user interests, behavioral conversion capabilities, and scenario adaptability, enabling the recommendation system to keenly capture the periodic and diverse product needs of maternal and infant users.

[0016] Through dynamic user lifecycle modeling and meticulous screening of similar users, the system can accurately identify a user's current stage of pregnancy or infant development, significantly improving the accuracy of matching specific needs across different stages. The recommendation process combines multiple signals, including collaborative filtering scores, user historical behavior scores, and stage matching scores, using a weighted hybrid model to comprehensively rank target products. This ensures that recommendations are adaptable to each stage, relevant to user preferences, and responsive to specific scenarios. In particular, during lifecycle transitions and emergencies, the system can quickly respond to changes in user behavior and deliver personalized products that meet current needs, significantly improving the real-time and accuracy of recommendations.

[0017] The proposed system structure and algorithmic model adapt to the dynamic, multi-stage nature of maternal and infant e-commerce users throughout their lifecycle, effectively avoiding outdated or irrelevant product recommendations. Practical application results show that the recommendation system's response speed, user satisfaction, and recommendation conversion rate significantly outperform existing general recommendation algorithms. This system can bring greater user stickiness and commercial value to maternal and infant platforms, demonstrating promising prospects for widespread application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flowchart of the method for dynamically recommending maternal and infant products based on user behavior proposed by the present invention; Figure 2 This is a user lifecycle state transition flow chart of the dynamic recommendation method for maternal and infant products based on user behavior proposed by the present invention; Figure 3This is a module structure diagram of the dynamic recommendation system for maternal and infant products based on user behavior proposed by the present invention; Figure 4 This is a schematic diagram of the system architecture of the dynamic recommendation system for maternal and infant products based on user behavior proposed by the present invention. DETAILED DESCRIPTION

[0019] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0020] refer to Figure 1-2 , a dynamic recommendation method for maternal and infant products based on user behavior, including the following steps: Collect users' multi-dimensional behavior data in real time on the client side, perform data cleaning and feature extraction on the multi-dimensional behavior data, and store it in a structured data format, providing an accurate and rich data foundation for subsequent life cycle modeling and recommendation algorithms; Dynamically establish the user's lifecycle status based on the user's input due date or baby's birth date and multi-dimensional behavioral data. This lifecycle status is divided using built-in pregnancy and baby growth stage rules. Finite state automata and Bayesian inference mechanisms are combined with behavioral data to correct and revise the lifecycle status based on specific user behaviors (such as purchasing certain stage-specific products or seeking emergency health advice). This effectively improves the intelligence level of lifecycle modeling and ensures the adaptability and foresight of subsequent recommendations. The collected multi-dimensional behavioral data is weighted according to the time interval between the behavior, the behavior type, and the current scenario label, using the time decay factor, behavior type weight, and scenario relevance parameter to calculate the behavioral feature weight. Specifically, for each behavior, the time decay factor is first applied based on the time interval between the behavior and the current moment to reduce the impact of historical behavior on current interest. Based on the behavior type (such as browsing, adding to cart, purchasing, etc.), the preset behavior type weight parameter is called to quantify the contribution of different behaviors to conversion. Combined with the current scenario label and product category, the scenario relevance matrix is ​​referenced to modify the behavior weight. Through this multiple weighting, we can obtain the behavioral feature weight that reflects the user's current interest preferences, behavior freshness, and scenario adaptability, achieving a refined expression of the behavior value. Based on the weights of behavioral features in the most recent period, behavior vectors are generated according to product categories. Combined with a stage synchronization adjustment factor, weighted cosine similarity is used to calculate the similarity between users and a reference user group. This filter selects similar user sets and introduces a stage synchronization adjustment factor to select only users in the same or adjacent life cycle stages as candidate similar users. This significantly improves the stage relevance and recommendation effectiveness of the similar user set. Based on the historical purchasing behavior of similar users and the weight of the user's own behavior, the collaborative filtering score and historical behavior score are calculated respectively. At the same time, the stage matching score is calculated according to the degree of overlap between the user's current life cycle stage and the product applicable stage. A weighted hybrid model is used to perform weighted summation of the collaborative filtering score, historical behavior score and stage matching score to generate a recommendation score for each target product. The product is sorted by score, abnormal products are excluded, and the final recommendation list is output. This method can fully reflect the stage needs and real-time behavior changes of maternal and infant users, and realize accurate, efficient and dynamic intelligent recommendation of maternal and infant products.

[0021] In this embodiment, the multidimensional behavioral data includes basic behaviors, implicit behaviors and scenario behaviors. Basic behaviors include browsing products, adding to cart, collecting, purchasing, and commenting. Implicit behaviors include page dwell time, scrolling depth, number of image magnifications, and click heat value. Scenario behaviors include user search keywords, consultation content, and browsing paths. The data cleaning includes filtering abnormal data, including browsing with a dwell time of <1 second, browsing with a scrolling depth of <10%, and repeated clicks of the same IP >10 times within one minute. The feature extraction uses an LSTM model to extract keywords from search terms and map them to a preset scenario tag library, such as "6 months" is mapped to the "complementary food addition period" and "eczema" is mapped to the "skin care period."

[0022] In this embodiment, the data correction step includes: When a user first registers, they enter the due date or the baby's birth date to generate the initial life cycle stage; Based on the passage of time, the current gestational age or baby's age is calculated and updated every 24 hours; Verification is performed based on user-specific behavior. When it is detected that a user purchases a product corresponding to a specific stage, the infant stage is corrected to the product applicable range. If the user purchases "stage 1 milk powder", the infant stage is confirmed to be 0-6 months; When the life cycle stage calculated based on time is inconsistent with the user's actual behavior data, a pop-up prompt is triggered to prompt the user to confirm whether the baby's current stage needs to be corrected. After the user confirms, the life cycle stage is updated through the state machine mechanism.

[0023] In this embodiment, the step of calculating the behavioral feature weight includes: We weight various types of user behavior data based on the time intervals between behaviors using a time-series decay function to obtain a time-series decay factor that reflects the "freshness" of the behavior. This time-series decay factor can effectively reduce the interference of historical behavior on current interest predictions, significantly enhancing the influence of recent behavior in recommendation results. Based on the user's current life cycle stage, the product category is assigned a scenario relevance parameter to obtain a scenario weight factor. The scenario weight factor is set at key life cycle nodes. For each key life cycle node, the system presets the scenario weight factor for each product category at that stage based on expert knowledge in the maternal and infant field and user behavior statistics. For example, in the "baby 4-6 months old" stage, the scenario weight factor of complementary food products is higher than that of newborn products. Through the above scenario relevance setting, the recommendation system can automatically perceive the core needs of the user at the current stage, significantly improving the stage adaptability of recommendations; The behavior type weight is obtained based on the statistics of the user's historical behavior data, specifically including: Collect massive amounts of user behavior data covering a certain period, including browsing, adding to cart, purchasing, favorites, and comments; For each behavior type, calculate the probability of a user ultimately converting to a purchase after completing the behavior. Using the conversion probability of browsing behavior as the benchmark value, calculate the conversion contribution coefficients of other behavior types relative to browsing behavior; The contribution coefficient is obtained by dividing the conversion probability of a certain behavior type by the conversion probability of browsing behavior. The normalization method is used to adjust the contribution of all behavior types to the range of 0 to 1 to obtain the behavior type weight; The normalization method uses "maximum normalization," which divides the contribution of each behavior by the maximum value to obtain the corresponding weight. This weight is then tested in actual business A / B testing. A / B testing involves randomly dividing users into two control groups within the same time period. Different weighting schemes are applied to each group of users, and their purchase conversion rates, click-through rates, and other indicators are compared to verify the consistency of the normalized weights with actual business results. The final behavior feature weight is obtained by multiplying the behavior type weight, the temporal decay factor, and the scenario weight factor. This behavior feature weight serves as an important input to the product interest score in the recommendation algorithm, ensuring that recommendation decisions take into account behavior freshness, behavior conversion value, and stage adaptability, effectively improving the relevance of system recommendations and user satisfaction.

[0024] In this embodiment, the calculation process of the timing attenuation factor specifically includes: Determine the time interval between a user's behavior and the current one; The weight window period corresponding to the query behavior type is used to indicate the length of time that the behavior can maintain the maximum weight influence after it occurs. The window period is set for different behavior types. For example, the window period for behaviors such as purchases and add-to-carts is longer, while the window period for behaviors such as browsing and commenting is relatively short. The weight window period is set by the system administrator based on user behavior statistics and business needs, and the unit is day. If the time interval between the occurrence of the behavior and the current weight window period is less than the current weight window period, the behavior's timing decay factor is set to 1. During this window period, the behavior's timing decay factor will not be decayed and the maximum weight will always be maintained. If the time interval between the occurrence of the behavior and the weight window period has exceeded, the temporal attenuation factor of the behavior is calculated as follows: Calculate the duration that the behavior occurs outside the window period. Subtract the weighted window period of the corresponding behavior type from the time interval between the behavior occurrence time and the current time to obtain the duration outside the window period. Based on the preset decay rate parameters of the behavior type, an exponential decay method is used. The specific calculation is to use the decay rate as the base and the duration exceeding the window period as the exponent to obtain a value less than 1 as the temporal decay factor of the behavior. The specific formula is: ; i represents the sequence number or category index of the behavior, which is used to distinguish different types of user behavior (such as browsing, adding to cart, purchasing, etc.). The window period corresponding to the i-th type of behavior refers to the length of time that the behavior can maintain the maximum weight influence after it occurs, usually in days or hours. Represents the decay rate parameter corresponding to the i-th type of behavior, which is used to control the decreasing speed of the behavior weight after the window period is exceeded. The larger the value, the faster the weight decreases. Represents the time interval from the occurrence of this behavior to the current moment, Represents the behavior of type i The timing decay factor, Represents the natural constant, which is the base of the exponential decay formula.

[0025] In this embodiment, the step of screening the similar user set includes: All behavioral data of target users within a recent preset period (e.g., the past 30 days) is categorized, counted, and aggregated by product category to form a behavioral feature vector based on each core maternity and baby product category. This feature vector can comprehensively reflect the user's interest preferences and behavioral activity in different product categories. The same method is used to generate behavioral feature vectors for all candidate users, and weighted cosine similarity is calculated with the target user's vector. Specifically, the system weights each dimension of the behavioral feature vector based on the weights or lifecycle-related parameters of each category. The similarity score between the target user and each candidate user is then calculated by calculating the cosine value of the angle between the weighted vectors. This method not only captures the overall similarity of user behavior patterns, but also highlights the importance of aligning key categories at the current lifecycle stage. Based on the similarity calculation results above, the top N users (e.g., N=50) with the highest similarity are selected from all candidate users as the similar user set of the target user. This set will be used for collaborative filtering and interest mining in the subsequent recommendation process, effectively improving the recommendation system's ability to match stage preferences and behavioral diversity, and optimizing the effect of personalized recommendations.

[0026] In this embodiment, the weighted cosine similarity specifically includes: In each category dimension of the behavioral feature vector, a stage-related weight is assigned to each product category based on the current life cycle stage of the target user and candidate user. This weight is determined by life cycle segmentation, category adaptability, and expert knowledge base. It aims to highlight the differences in users' attention to products in different categories at different stages. The method for setting the relevant weights of the stages is: Based on the life cycle segmentation rules, users are divided into specific stage intervals, such as early pregnancy, mid-pregnancy, late pregnancy, neonatal period, infants 6-12 months old, etc.

[0027] Based on the preset category adaptability matrix, determine the matching relationship between each stage and different product categories. For example, increase the weight of folic acid products in early pregnancy and increase the weight of complementary food products for infants aged 6-12 months.

[0028] Combined with the recommended priorities for each stage given by maternal and infant expert knowledge base, the above category matching relationships are mapped to stage-related weights. This can be achieved through a table lookup method, that is, presetting weight values ​​for each stage-category combination.

[0029] The relevant weights of the above stages range from 0 to 1. The larger the value, the higher the user's attention to the category at that stage.

[0030] Before calculating cosine similarity, the behavior vectors of the target user and candidate users are weighted and normalized to obtain stage-weighted behavior vectors. The behavioral feature values ​​of each dimension are multiplied by the corresponding stage-related weight, and the vector is normalized as a whole to enable equivalent comparison of different users' interest characteristics in various categories. Specifically, the square sum of the values ​​of each dimension of the weighted vector is calculated, and the square root is taken to obtain the vector norm (i.e., L2 norm). The weighted feature value of each dimension is then divided by this norm, and the data of all dimensions is mapped to the range of 0-1. This process can improve the stage sensitivity of the behavior vector and make the similarity measurement more accurately reflect the real needs of users in the current stage. A phase synchronization adjustment factor is introduced to characterize whether the target user and the candidate user have overlapping active periods in the same category. Specifically, the system detects the high-activity time windows of the target user and the candidate user in a certain category. If the active periods of the two overlap, the set synchronization factor is introduced into the similarity calculation to enhance the similarity score of the category; otherwise, the similarity score of the category is reduced or not included. This can effectively enhance the influence of the time alignment of key user behaviors and avoid inflated similarity due to interest transfer or life cycle mismatch. In the final similarity score, the weighted cosine value is multiplied by the stage synchronization adjustment factor to obtain the similarity calculation result, which serves as the final inter-user similarity score. Through this method, the system can accurately characterize the degree of interest matching between users across multiple dimensions such as behavioral patterns, lifecycle stages, and behavioral timing, significantly improving the stage adaptability and personalization capabilities of the recommendation system in the maternal and infant scenario.

[0031] In this embodiment, the step of outputting the recommendation list specifically includes: For each product to be recommended, we traverse the previously filtered similar user set and count how many users in the similar user set have purchased the product to be recommended. The number of users who actually purchased the product to be recommended is divided by the total number of similar users in the set to obtain the collaborative filtering score of the product to be recommended. This score can reflect the popularity of the product among users with similar interests and help to mine community recommendation signals. For each product to be recommended, we summarize all the target user's recent behaviors related to the product, multiply them by their respective behavioral feature weights, add up the weighted results, and normalize them to obtain the historical behavior score of the product to be recommended. This score can fully reflect the user's individual interests, preferences, and recent activity. Obtain the time interval of the current user's life cycle stage and the product's applicable life cycle interval, calculate the overlap duration of the two time intervals, and then divide the overlap duration by the total duration of the product's applicable interval to obtain the stage overlap ratio. Then, multiply the stage overlap ratio by the preset stage matching correction coefficient to finally obtain the stage matching score. This score can reflect the product's adaptability to the user's current growth stage and improve the accuracy and professionalism of recommendations; For each product to be recommended, the collaborative filtering score, historical behavior score, and stage matching score are weighted and summed according to the preset weights (for example, the weight ratio is 0.4:0.3:0.3) to form the final comprehensive recommendation score for the product to be recommended. The products are then sorted from high to low based on the comprehensive recommendation scores to generate a recommendation list. The products in the recommendation list are checked one by one, and products that have been purchased by the user are eliminated to avoid duplicate recommendations. At the same time, products with prices higher than twice the average of the user's historical purchase prices are eliminated to filter out abnormally high-priced products to ensure the quality of recommendations. Jaccard similarity is further used to compare the user's historical purchase price set with the set of product prices to be recommended. When the price deviation is greater than the set threshold, the product will not be recommended, and the final recommendation list will be generated.

[0032] refer to Figure 3-4 , a dynamic recommendation system for maternal and infant products based on user behavior, including the following modules: The multi-dimensional behavior data collection module is used to collect multi-dimensional behavior data of users, including basic behavior, implicit behavior and scenario behavior data, from the client in real time, and perform data cleaning and feature extraction on the collected multi-dimensional behavior data, and store it in a structured data format; A user lifecycle modeling module is used to establish a dynamic state machine of the user lifecycle based on the due date or baby birth date information input by the user and the behavioral data uploaded by the multi-dimensional behavioral data acquisition module, and to correct the lifecycle state and switch stages using finite state automata and Bayesian inference rules; The behavior feature weight calculation module is used to calculate the behavior feature weight by weighting the collected multi-dimensional behavior data according to the time interval of the behavior, the behavior type and the current scene label using the time series attenuation factor, the behavior type weight and the scene relevance parameter respectively; The dynamic similarity calculation module is used to aggregate the weights of users' behavioral features by product category based on a set time period to construct a behavior vector. This vector is then combined with a stage synchronization adjustment factor to analyze the similarity of the target user's behavior patterns with other users in the same or adjacent lifecycle stages using weighted cosine similarity, thereby screening out a set of similar users. The recommendation result generation module is used to integrate three types of scoring results: collaborative filtering score, user historical behavior score and stage matching score, and weighted summary according to the preset hybrid weight model to complete the comprehensive recommendation score sorting of the target products, sort them by score, exclude abnormal products, and output the final recommendation list.

[0033] Example 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to a shopping software. Take user A as an example. His baby is currently 5 months old. His recent key behaviors include: browsing the "6-month complementary food" page 3 days ago (staying on the page for 120 seconds), purchasing "high-iron rice noodles" 1 week ago, and recently searching for the keyword "baby's first complementary food".

[0034] Based on the baby's birth date and recent behavior filled in by User A, the system infers that the baby is in the "5-month-old infant" stage. Based on the stage-product matching table, the system significantly improves the matching degree of all "complementary food" products, giving priority to the complementary food needs of the corresponding stage.

[0035] Browsing behavior for "six-month solid food" was weighted by combining page dwell time, temporal decay factor, and scenario weight, resulting in a higher score. The weight of adding "high-speed iron rice noodles" to the shopping cart was further increased due to the combined effect of behavior type weight and temporal decay. Searches for "baby's first solid food" were identified as highly relevant interest tags in the current scenario, significantly increasing the weight of products related to solid food.

[0036] The system searches the platform for all parents of infants aged 5-6 months and uses weighted cosine similarity calculations based on behavioral characteristics to identify the top 50 users with interests most similar to User A. This group recently reported a 60% purchase rate for "supplementary feeding spoons." They also frequently purchase products like "grinding bowls" and "six-month rice cereal" during the feeding period.

[0037] Taking "supplementary food spoon" as an example, the system combines the collaborative filtering score (60% of similar users purchase it), the historical behavior score (the cumulative score of user A's recent supplementary food-related behavior is 0.45), and the stage matching degree (the current 5-month age and the "7-month supplementary food" product compatibility is 0.9) for weighted mixing, and the comprehensive recommendation score Score = 0.4×0.6+0.3×0.45+0.3×0.9=0.645.

[0038] Finally, a recommendation list is generated, with products such as "food spoon", "6-month rice noodles", and "grinding bowl" being preferentially output and pushed to the user end via WebSocket. The products with higher rankings are the most suitable products for users at the current stage.

[0039] Table 1 Comparison of recommendation system indicators

[0040] As shown in Table 1, the dynamic recommendation system for maternal and infant products of the present invention significantly outperforms existing solutions in many core indicators. Recommendation accuracy (NDCG@10) increased from 0.54 to 0.74, a 37% improvement. In key application scenarios such as fever and the complementary feeding period, the system's scenario hit rate exceeded 0.85, significantly higher than the control group. The 95th percentile of the system's recommendation response speed dropped to 8.2 seconds, meeting and exceeding the industry standard of less than 10 seconds, significantly improving the user's instant experience. In terms of commercial effectiveness, the platform's actual operating data shows that user conversion rate increased by 22.5%, average order value increased by 14.5%, the proportion of irrelevant recommendations decreased significantly, and the compatibility of recommendation results with the infant's growth stage increased to 93%, greatly optimizing the user experience and platform operating efficiency.

[0041] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A dynamic recommendation method for maternal and infant products based on user behavior, characterized by: The steps include: Collect users' multi-dimensional behavior data in real time on the client side, perform data cleaning and feature extraction on the multi-dimensional behavior data, and store it in a structured data format to generate structured multi-dimensional behavior data; Dynamically establish the user lifecycle status based on the user's input due date or baby's birth date and multi-dimensional behavioral data, and perform data correction through a combination of finite state machine and Bayesian inference mechanism; For structured multidimensional behavior data, we use the time interval, behavior type, and current scene label to weight the behavior feature data using the time series decay factor, behavior type weight, and scene relevance parameter to calculate the behavior feature weight. Based on the weights of behavioral features in the most recent period, a behavioral vector is generated according to product category. Combined with the stage synchronization adjustment factor, weighted cosine similarity is used to calculate the similarity between the user and the reference user group to screen similar user sets; Based on the historical purchasing behavior of similar users and the user's own behavior, the collaborative filtering score, historical behavior score and stage matching score are calculated respectively. A weighted hybrid model is used to weight the collaborative filtering score, historical behavior score and stage matching score to generate a recommendation score for each target product. The products are then sorted by the recommendation score, abnormal products are excluded, and the final recommendation list is output.

2. The method for dynamically recommending maternal and infant products based on user behavior according to claim 1, characterized in that: The multi-dimensional behavioral data includes basic behaviors, implicit behaviors and scenario behaviors. Basic behaviors include browsing products, adding to cart, collecting, purchasing, and commenting. Implicit behaviors include page dwell time, scrolling depth, number of image magnifications, and click heat value. Scenario behaviors include user search keywords, consulting content, and browsing paths. The data cleaning includes filtering abnormal data, including browsing with a dwell time of <1 second, browsing with a scrolling depth of <10%, and repeated clicks of the same IP >10 times within one minute. The feature extraction uses an LSTM model to extract keywords from search terms and map them to a preset scenario tag library.

3. The method for dynamically recommending maternal and infant products based on user behavior according to claim 1, characterized in that: The data correction step comprises: When a user first registers, they enter the due date or the baby's birth date to generate the initial life cycle stage; Based on the passage of time, the current gestational age or baby's age is calculated and updated every 24 hours; Verify based on user-specific behavior. When it is detected that a user purchases a product corresponding to a specific stage, the infant stage is corrected to the product's applicable range. When the life cycle stage calculated based on time is inconsistent with the user's actual behavior data, a pop-up prompt is triggered to prompt the user to confirm whether the baby's current stage needs to be corrected. After the user confirms, the life cycle stage is updated through the state machine mechanism.

4. The method for dynamically recommending maternal and infant products based on user behavior according to claim 1, characterized in that: The step of calculating the behavioral feature weight comprises: The user's various behavioral data are weighted according to the time interval between behaviors using a time series decay function to obtain a time series decay factor. According to the current life cycle stage of the user, the product category is assigned a scenario relevance parameter to obtain a scenario weight factor, which is set by the key nodes of the life cycle; The behavior type weight is obtained based on the statistics of the user's historical behavior data, specifically including: Collect massive amounts of user behavior data covering a certain period, including browsing, adding to cart, purchasing, favorites, and comments; For each behavior type, calculate the probability of a user ultimately converting to a purchase after completing the behavior. Using the conversion probability of browsing behavior as the benchmark value, calculate the conversion contribution coefficients of other behavior types relative to browsing behavior; Use normalization to adjust the contribution of all behavior types to the range of 0-1, and conduct A / B testing in actual business. The behavior type weight, temporal decay factor, and scenario weight factor are multiplied together to obtain the final behavior feature weight.

5. The method for dynamically recommending maternal and infant products based on user behavior according to claim 4, characterized in that: The calculation process of the timing attenuation factor specifically includes: Determine the time interval between a user's behavior and the current one; Query the weight window period corresponding to the behavior type; If the time interval between the occurrence of the behavior and the current weight window period is less than the current weight window period, the behavior's timing decay factor is set to 1. During this window period, the behavior's timing decay factor will not be decayed and the maximum weight will always be maintained. If the time interval between the occurrence of the behavior and the weight window period has exceeded, the temporal attenuation factor of the behavior is calculated as follows: Calculate the duration of time the behavior occurs outside the window period; According to the preset decay rate parameters of the behavior type, an exponential decay method is adopted. The specific calculation is to use the decay rate as the base, take the duration exceeding the window period as the exponent, and obtain a value less than 1 as the temporal decay factor of the behavior.

6. The method for dynamically recommending maternal and infant products based on user behavior according to claim 1, characterized in that: The step of screening the similar user set includes: All behavioral data of target users in the most recent preset period is classified, counted, and aggregated by product category to form a behavioral feature vector based on each core maternity and baby product category. Generate behavioral feature vectors for all candidate users in the same way and calculate weighted cosine similarity with the target user's vector; According to the weighted cosine similarity, the top N users with similarity are selected from all candidate users as the similar user set of the target user.

7. The method for dynamically recommending maternal and infant products based on user behavior according to claim 1, characterized in that: The calculation process of the weighted cosine similarity specifically includes: In each category dimension of the behavioral feature vector, assign stage-related weights to each product category based on the current life cycle stage of the target user and candidate user; Before calculating the cosine similarity, the behavior vectors of the target user and the candidate user are weighted and normalized to obtain the stage-weighted behavior vectors; Introducing a phase synchronization adjustment factor, only when the target user and the candidate user's category activity period overlap, will the set similarity score be given; In the final similarity score, the weighted cosine value is multiplied by the phase synchronization adjustment factor to obtain the similarity calculation result.

8. The method for dynamically recommending maternal and infant products based on user behavior according to claim 1, characterized in that: The step of outputting the recommendation list specifically includes: For each product to be recommended, the total number of users in the similar user set who have purchased the product to be recommended is counted. The number of users who actually purchased the product to be recommended is divided by the total number of similar user sets to obtain the collaborative filtering score of the product to be recommended. For each product to be recommended, summarize all the target user's recent behaviors related to the product to be recommended, multiply them by their respective behavioral feature weights, add up the weighted results and normalize them to obtain the historical behavior score of the product to be recommended; Obtain the time interval of the current user's life cycle stage and the product's applicable life cycle interval, calculate the overlap duration of the two time intervals, divide the overlap duration by the total duration of the product's applicable interval to obtain the stage overlap ratio, and multiply the stage overlap ratio by the preset stage matching correction coefficient to obtain the final stage matching score; For each product to be recommended, the collaborative filtering score, historical behavior score, and stage matching score are weighted and summed according to the preset weights to serve as the final comprehensive recommendation score for the product to be recommended. The products are then sorted from high to low based on the comprehensive recommendation scores to generate a recommendation list. The products in the recommendation list are checked one by one, and products that have been purchased by the user are eliminated. At the same time, products with prices higher than twice the average price of the user's historical purchases are eliminated. Jaccard similarity is further used to compare the user's historical purchase price set with the set of product prices to be recommended. When the price deviation is greater than the set threshold, the product is not recommended, and the final recommendation list is generated.

9. A dynamic recommendation system for maternal and infant products based on user behavior, which implements the dynamic recommendation method for maternal and infant products based on user behavior according to any one of claims 1 to 8, characterized in that: Includes the following modules: The multi-dimensional behavior data collection module is used to collect multi-dimensional behavior data of users, including basic behavior, implicit behavior and scenario behavior data, from the client in real time, and perform data cleaning and feature extraction on the collected multi-dimensional behavior data, and store the multi-dimensional behavior data in a structured data format; A user lifecycle modeling module is used to establish a dynamic state machine of the user lifecycle based on the due date or baby birth date information input by the user and the behavioral data uploaded by the multi-dimensional behavioral data acquisition module, and to correct the lifecycle state and switch stages using finite state automata and Bayesian inference rules; The behavior feature weight calculation module is used to calculate the behavior feature weight by weighting the collected multi-dimensional behavior data according to the time interval of the behavior, the behavior type and the current scene label using the time series attenuation factor, the behavior type weight and the scene relevance parameter respectively; The dynamic similarity calculation module is used to aggregate the weights of users' behavioral features by product category based on a set time period to construct a behavior vector. This vector is then combined with a stage synchronization adjustment factor to analyze the similarity of the target user's behavior patterns with other users in the same or adjacent lifecycle stages using weighted cosine similarity, thereby screening out a set of similar users. The recommendation result generation module is used to integrate three types of scoring results: collaborative filtering score, user historical behavior score and stage matching score, and weighted summary according to the preset hybrid weight model to complete the comprehensive recommendation score sorting of the target products, sort them by score, exclude abnormal products, and output the final recommendation list.

Images