Large-scale commodity collaborative filtering recommendation method and system based on hybrid recommendation algorithm

Abstract

The application discloses a large-scale commodity collaborative filtering recommendation method and system based on a hybrid recommendation algorithm, and relates to the technical field of artificial intelligence and recommendation systems. The method comprises the following steps: collecting multi-dimensional data of users; constructing a virtual person survival scene based on the multi-dimensional data; calculating the estimated remaining days of goods based on virtual inventory and shopping habits, and triggering a consumable recommendation when the number is lower than a threshold; detecting changes in residence based on migration trajectories, and triggering a scene adaptive recommendation; integrating virtual survival scene features into collaborative filtering algorithms for enhanced recommendation; and generating a final recommendation list by integrating multiple recommendation results. The application generates regional demand forecasts by aggregating virtual survival scene data of each user, provides replenishment suggestions for merchants, and realizes bilateral collaborative optimization of supply and demand. The application solves the technical problems of the existing recommendation system, such as lack of life perception, inability to predict product consumption and user migration, and improves the accuracy, timeliness and user experience of the recommendation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and recommendation systems, and in particular to a large-scale collaborative filtering recommendation method and system for goods based on hybrid recommendation algorithms. Background Technology

[0002] With the rapid development of e-commerce and mobile internet, personalized recommendation systems have become a core tool for e-commerce platforms to improve user experience and conversion rates. Currently, mainstream recommendation technologies mainly include collaborative filtering-based recommendation, content-based recommendation, deep learning-based recommendation, and hybrid recommendation.

[0003] Several related recommendation technologies have been disclosed in Chinese patent literature. For example, CN108876508A discloses a hybrid recommendation method based on Hadoop and Spark frameworks. Under the Hadoop framework, a user-based collaborative filtering algorithm is used to perform offline analysis on massive amounts of data to generate an offline recommendation list. Under the Spark framework, an item-based collaborative filtering algorithm is used to generate a real-time recommendation list, and then the two are merged into a final recommendation list. CN110362755A discloses a recommendation method based on a hybrid algorithm of item collaborative filtering and association rules. It obtains item association rules by constructing a user behavior history dataset, constructs a user-item rating matrix to calculate item similarity, and simultaneously constructs an item content attribute matrix to calculate similarity based on item attributes. CN114461922A discloses a graph-based hybrid message passing mechanism for collaborative filtering information recommendation systems. It decouples the user-item information bipartite heterogeneous graph into two homogeneous graphs (user and item), and combines this with an implicit message passing method to form a hybrid message passing mechanism. CN113761393B discloses a product collaborative recommendation method, which obtains interaction behavior data of multiple behavior types corresponding to products, and calculates the weighted similarity by counting the total number of users of each behavior type between pairs of products.

[0004] In the field of virtual human technology, CN120013641A discloses an AI virtual human product matching optimization method and system. This method uses deep learning algorithms to extract features from user data and perform cluster analysis to construct user profiles, and dynamically adjusts the matching strategy based on an adaptive deep reinforcement learning model. CN119250863A discloses a customer lifecycle management system and method based on multi-channel integration, which predicts customer behavior by creating digital twins of customers. Furthermore, there are also behavior recommendation devices based on digital twins that output a prediction parameter set by inputting behavior sequences and monitoring parameter sets.

[0005] While the aforementioned technologies have achieved personalized recommendations to some extent, they still have the following significant shortcomings:

[0006] (i) Lack of modeling for all aspects of users' lives

[0007] Current recommendation systems primarily rely on user browsing, clicking, and purchasing behavior data from e-commerce platforms, failing to incorporate users' offline life scenarios into their modeling, resulting in a lack of real-life awareness in the recommendations. Users' true needs often stem from actual daily consumption and changes in context, signals that current technologies cannot capture.

[0008] (ii) Lack of ability to predict the consumption of goods

[0009] Current technologies cannot infer the usage progress and remaining lifespan of items based on information such as users' shopping intervals and usage habits. For example, after a user purchases a bottle of shower gel, the system cannot determine when it will run out, thus missing the opportunity for accurate restocking recommendations. Although some technologies involve IoT sensors to monitor the consumption of daily necessities, these methods rely on hardware deployment, are costly, lack versatility, and cannot cover all product types.

[0010] (iii) Lack of awareness and utilization of user migration behavior

[0011] When users relocate, their lifestyles and consumption needs change significantly, but existing recommendation systems cannot detect this change and adjust their recommendation strategies accordingly. Current user-trajectory-based service recommendations are mainly used for scenarios such as route planning, and do not cover scenario adaptation for product recommendations.

[0012] (iv) Staticization of user profiles

[0013] Existing user profiles are mostly updated statically or quasi-statically, lacking the ability to continuously track and simulate the dynamic evolution of users' life status, and thus failing to reflect the dynamic changes in users' life scenarios.

[0014] (v) The recommendation system is a one-way information flow

[0015] Existing recommendation systems are one-way information flows from users to products. The system cannot feed back the aggregated user demand signals to merchants and the supply chain, and cannot achieve collaborative optimization between the supply and demand sides.

[0016] Therefore, we propose a large-scale collaborative filtering recommendation method and system for goods based on a hybrid recommendation algorithm. Summary of the Invention

[0017] The purpose of this invention is to address the shortcomings of existing technologies by providing a large-scale collaborative filtering recommendation method and system for goods based on hybrid recommendation algorithms, thereby solving the technical problems mentioned in the background section.

[0018] To achieve the above objectives, the present invention provides the following technical solution:

[0019] A large-scale collaborative filtering recommendation method for goods based on hybrid recommendation algorithms includes the following steps:

[0020] S1: Collect multi-dimensional data of users, including shopping data, address data, behavior data and spatiotemporal data;

[0021] S2: Construct a virtual character survival scenario based on the multi-dimensional data. The virtual character survival scenario includes at least: the virtual character's spatial location information, virtual inventory information, shopping habit information, and migration trajectory information.

[0022] S3: Based on the virtual inventory information and the shopping habit information, calculate the estimated remaining usage days for each item;

[0023] S4: When the estimated remaining usage days are lower than a preset threshold, a consumption-triggered recommendation is triggered.

[0024] S5: Detect residence change events based on the migration trajectory information. When a residence change is detected, trigger scene change-triggered recommendations.

[0025] S6: The features of the virtual character's survival scene are used as an enhanced dimension of the user profile and integrated into the collaborative filtering algorithm for enhanced collaborative filtering recommendations.

[0026] S7: Combine the results of the consumption-triggered recommendations, the scene change-triggered recommendations, and the collaborative filtering-enhanced recommendations to generate the final recommendation list.

[0027] Preferably, the method for constructing the virtual inventory information includes:

[0028] When a user purchases a product, the initial stock of that product in the virtual inventory is initialized based on the product specifications and category attributes.

[0029] The user's average daily consumption rate is estimated based on the user's historical purchase intervals for this product category, and then corrected by combining it with the average consumption rate of similar users to establish a personalized item consumption curve.

[0030] Based on the initial inventory and the average daily consumption rate, the estimated remaining days for each item are calculated in real time.

[0031] Preferably, the estimated remaining usage days are obtained by dividing the current remaining quantity of the item by the user's average daily consumption rate for that category; the average daily consumption rate is determined as follows: for users with existing historical purchase records, the average consumption per unit time is calculated based on the quantity of items purchased multiple times within a preset historical time window and the corresponding time points; for users without historical purchase records, the average consumption rate of similar users in that category is used as the initial value, and is gradually updated to personalized parameters after the user generates subsequent purchase records.

[0032] Preferably, the method for constructing the migration trajectory information includes:

[0033] Continuously monitor changes in users' delivery addresses. When the new address is inconsistent with the city or district information in the historical address, record the migration event. The migration event includes migration time, place of origin, place of destination, and migration type.

[0034] Analyze the changes in consumption characteristics before and after migration, and establish a mapping relationship between migration and consumption;

[0035] Predict users' future migration probability based on historical migration patterns and external data.

[0036] Preferably, the scene change-triggered recommendations include:

[0037] Migration-triggered recommendations: When a user's place of residence is detected to have changed, the city feature vector of the destination is obtained, the city feature vector is matched with the user's historical consumption characteristics, the set of new demand categories that the user may generate in the new city is calculated, and a migration-triggered recommendation list is generated.

[0038] Seasonal recommendations: Recommend seasonal products based on the climate and seasonal changes in the location of the virtual character;

[0039] Recommendations triggered by changes in life stage: Based on changes in the user's consumption category structure, we infer changes in life stage and recommend suitable products for the corresponding life stage.

[0040] Preferably, in the collaborative filtering enhanced recommendation, the user's virtual survival scene features are encoded into scene feature vectors, and the user's historical interaction behavior is encoded into interaction feature vectors, which are concatenated to form a comprehensive user feature vector. When calculating the similarity between users, the scene similarity of two users on the scene feature vector and the interaction similarity on the interaction feature vector are calculated separately. The scene similarity and the interaction similarity are weighted and summed with preset weights to obtain the comprehensive user similarity. Based on the comprehensive user similarity, neighboring users are determined, the interaction of the neighboring users with the target product is aggregated, and the current user's preference score for the target product is predicted.

[0041] Preferably, the following steps are also included:

[0042] After standardizing the virtual survival scenario data of each user, the data is aggregated in the spatiotemporal dimension to generate regional and category-level demand forecasts.

[0043] Based on the demand forecast, replenishment suggestions are generated for merchants, including replenishment items, suggested replenishment quantity, suggested replenishment time, and suggested replenishment location;

[0044] The replenishment suggestion will be pushed to the corresponding merchant terminal;

[0045] Based on the aforementioned demand forecast, the platform's inventory distribution and logistics scheduling are optimized.

[0046] Preferably, the step of aggregation in the spatiotemporal dimension includes:

[0047] Aggregation by time dimension: Merge and statistically analyze the estimated expiration dates of the same product or category for each user to generate a time series demand forecast curve;

[0048] Aggregation by spatial dimension: Linking the geographical location information of each user with demand forecasts to generate a regional demand heat map;

[0049] Aggregation by migration dimension: Correlate user migration predictions with new demand at destinations to generate migration-driven regional demand alerts.

[0050] Preferably, the recommended replenishment quantity is determined by multiplying the predicted demand by a safety factor, which is greater than one and its specific value is dynamically adjusted based on the historical demand forecast error and supply stability of the product category; the recommended replenishment quantity is allocated in the replenishment recommendation according to the demand forecast ratio of each sub-region.

[0051] A large-scale collaborative filtering recommendation system for goods based on hybrid recommendation algorithms includes:

[0052] The data acquisition module is used to collect multi-dimensional data from users, including shopping data, address data, behavioral data, and spatiotemporal data.

[0053] The virtual scene construction module is used to construct a virtual character survival scene based on the multi-dimensional data. The virtual character survival scene includes at least the virtual character's spatial location information, virtual inventory information, shopping habit information, and migration trajectory information.

[0054] The item consumption prediction module is used to calculate the estimated remaining number of days of use for each item based on the virtual inventory information and the shopping habit information;

[0055] The scene change detection module is used to detect residence change events and life stage change events based on the migration trajectory information;

[0056] The recommendation generation module is used to generate consumption-triggered recommendations based on the estimated remaining days of use, and to generate scenario-change-triggered recommendations based on the residence change event and life stage change event.

[0057] The collaborative filtering enhancement module is used to incorporate the features of the virtual character's survival scene as an enhanced dimension of the user profile into the collaborative filtering algorithm for collaborative filtering enhancement recommendation, and to generate the final recommendation list by combining the results of consumption-triggered recommendation, scene change-triggered recommendation and collaborative filtering enhancement recommendation.

[0058] The demand aggregation module is used to aggregate virtual survival scenario data of various users in the spatiotemporal dimension to generate regional and category-level demand forecasts.

[0059] The replenishment suggestion generation module is used to generate replenishment suggestions for merchants based on the demand forecast.

[0060] The supply and demand matching optimization module is used to optimize the inventory distribution and logistics scheduling on the platform side based on the demand forecast.

[0061] The feedback optimization module is used to continuously optimize the model parameters of each module based on user and merchant feedback.

[0062] The beneficial effects of this invention are as follows:

[0063] 1. This invention innovatively expands the modeling perspective of recommendation systems from users' internal platform behavioral data to their entire offline life scenarios. By constructing a four-dimensional virtual avatar survival scenario encompassing spatial positioning, item inventory, shopping habits, and migration trajectories, the recommendation system gains the ability to "perceive the user's life status." Traditional recommendation systems can only answer "what the user might like," while this invention can answer "what the user needs in life," achieving a fundamental leap in recommendation logic from preference matching to life scenario deduction. The resulting technical effects are: the recommendation system no longer passively waits for users' browsing or search behavior, but actively perceives the user's consumption progress and scenario changes in life, completing demand prediction before the user actively expresses their needs. Recommendation timing is more accurate, recommended content is more closely aligned with the user's real life status, and user click-through rates and conversion rates are significantly improved.

[0064] 2. This invention does not rely on any IoT sensors or other hardware devices. It establishes personalized item consumption curves and accurately predicts the remaining usage days of each item simply by analyzing the historical purchase intervals and quantities of similar products purchased by users. This technical solution is far less expensive than sensor-dependent IoT replenishment solutions and can cover all product types, exhibiting good universality and scalability. When the estimated remaining days of an item are lower than a preset threshold, the system automatically triggers consumable recommendations, realizing a shift from a passive to an active recommendation paradigm from "users remembering to buy" to "system reminding them to buy." The resulting technical effects are: a significantly shortened repurchase cycle for daily consumables, eliminating the awkward situation of users only remembering to buy after running out of items, and effectively improving the platform's overall repurchase rate and user lifetime value.

[0065] 3. This invention continuously monitors changes in users' delivery addresses to accurately identify user relocation events and their types (such as graduation, job relocation, returning home, property purchase, etc.). Based on the matching of the characteristics of the destination city with the user's historical consumption characteristics, it automatically generates a suitable recommendation list for the relocation scenario. This technical solution solves the technical problem that existing recommendation systems cannot perceive drastic changes in users' life scenarios and that recommended content is out of touch with actual needs. The resulting technical effect is that when a user relocates, the system can accurately recommend products that the user actually needs in the new life scenario (such as storage supplies after moving, products adapted to the new climate, and specific categories of products needed in the new life stage). The matching degree between recommended content and the user's current life status is greatly improved, and the user's relocation adaptation experience is significantly improved.

[0066] 4. This invention, through the construction of virtual survival scenario features, enables new users to initially construct a virtual survival scenario using limited information such as their shipping address and the category of their first purchase, even with only a small amount of interaction data. This allows them to establish connections with established users who have similar survival scenarios, thereby generating effective recommendations. This technical solution overcomes the limitation of traditional collaborative filtering algorithms that rely entirely on user interaction data, transforming the cold start problem from "insufficient data for recommendations" to "recommendations based on scenario matching." The resulting technical effects are: significantly improved quality of first-screen recommendations for new users, effective improvement in new user retention and initial conversion rates, and a substantial reduction in the gap between the recommendation performance in cold start scenarios and that of established users.

[0067] 5. This invention not only utilizes virtual survival scenarios for personalized recommendations on the user side, but also aggregates massive amounts of virtual survival scenario data from users to generate regional and category-level demand predictions, which are then pushed back to merchants to form replenishment suggestions. This technical solution breaks the existing one-way information flow pattern of "user to product" in recommendation systems, constructing a complete closed loop of "user demand perception → platform aggregation and prediction → merchant precise replenishment → product pre-deployment → improved user experience." The resulting technical effects are: merchants gain actionable and precise replenishment decision support, improving inventory turnover, reducing stockout rates, and lowering inventory holding costs; the platform achieves intelligent warehousing and logistics optimization, improving delivery timeliness and reducing logistics costs; and users gain richer choices and faster delivery services. The value of all three parties is simultaneously enhanced, transforming the recommendation system from a "cost center" to a "value center," significantly improving the system's commercial sustainability and social benefits.

[0068] 6. This invention collects implicit and explicit user feedback on recommendation results, as well as deviation data between the merchant's adoption rate of replenishment suggestions and actual demand, through a feedback optimization module. It periodically performs incremental training and parameter tuning on each prediction and recommendation model. This technical solution enables the system to achieve closed-loop continuous evolution. As the running time increases, the accuracy of each prediction model continuously improves, the recommendation effect is continuously enhanced, and the system's adaptability, robustness, and long-term effectiveness are effectively guaranteed. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of the large-scale collaborative filtering recommendation method for goods based on a hybrid recommendation algorithm according to the present invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Example

[0072] This embodiment discloses a large-scale collaborative filtering recommendation method for goods based on a hybrid recommendation algorithm.

[0073] S1: The system first collects multi-dimensional user data through the data acquisition module. Specifically, the system obtains user shopping data through e-commerce platform APIs, address data through user account information, behavioral data through event tracking logs, and spatiotemporal data through IP location and LBS services.

[0074] S2: The system then uses the virtual scene construction module to construct a virtual character survival scenario for each user based on the aforementioned multi-dimensional data. Taking User A as an example, the system determines that their delivery address is Zhongguancun Street, Haidian District, Beijing, and identifies the community type as a science park-type residential area; the system maintains virtual inventory based on their historical purchase records, recording the initial stock and consumption progress of items such as shampoo, shower gel, and tissues; the system extracts their shopping habits, including the approximately 110-day purchase cycle for shampoo and their loyalty to specific brands; the system tracks their address change history, recording their migration event from Xueyuan Road address to Zhongguancun Street address.

[0075] S3: The system uses the item consumption prediction module to calculate the estimated remaining days of use for each item based on virtual inventory information and shopping habit information. Taking shampoo in user A's current virtual inventory as an example, the system estimates the average daily consumption based on the time interval and quantity of shampoo purchased twice, and calculates the estimated remaining days based on the current remaining quantity.

[0076] S4: When the estimated remaining days are lower than a preset threshold (seven days in this embodiment), the system triggers a consumption-triggered recommendation through the recommendation generation module. The system selects a recommendation strategy based on the user's shopping habits, prioritizing repeat purchase recommendations of the same product, while also searching for new or upgraded versions of the brand as upgrade alternatives, and searching for highly-rated products in the same category as competitor recommendations.

[0077] S5: The system detects residence change events based on migration trajectory information through the scene change detection module. When a user's primary delivery address changes from Haidian District, Beijing to Nanshan District, Shenzhen, Guangdong Province, the system records the migration event and identifies the migration type as workplace relocation.

[0078] When a change of residence is detected, the system triggers scenario-driven recommendations through the recommendation generation module. The system obtains the urban characteristics of Nanshan District, Shenzhen, matches them with the user's historical consumption characteristics, predicts that the user may need sun protection products, dehumidifiers, light summer clothing, as well as storage supplies and cleaning tools after moving, and generates a migration-triggered recommendation list.

[0079] S6: The system incorporates features of virtual character survival scenarios as enhanced dimensions into the collaborative filtering algorithm through a collaborative filtering enhancement module. When calculating user similarity, the system considers both the similarity of purchasing behavior and the similarity of survival scenarios.

[0080] S7: The system integrates three results—consumption-triggered recommendation, scenario-change-triggered recommendation, and collaborative filtering-enhanced recommendation—and generates a final recommendation list through multi-objective optimization before pushing it to the user's terminal.

[0081] Preferably, the system maintains a virtual inventory for each user, and the construction process is as follows:

[0082] When a user purchases a product, the system first initializes the initial inventory of that item in the virtual inventory based on the product specifications and category attributes. For example, if a user purchases a 500ml bottle of shower gel, the system creates a record for that shower gel in the virtual inventory, with an initial inventory of 500ml.

[0083] The system then calculates the user's average daily consumption rate for this product category based on the user's historical purchase intervals. Specifically, the system retrieves all purchase records of this product category within a certain time window, calculates the time interval between two consecutive purchases, and uses the statistical value of multiple purchase intervals to characterize the user's consumption rate for this product category. For example, if the time interval between the user's two previous purchases of shower gel was approximately 100 days, the system would use an average daily consumption of approximately 5ml as the user's shower gel consumption rate parameter.

[0084] When a user lacks historical purchase data for a certain product category, the system uses the average consumption rate of similar users as the initial parameter. Specifically, the system retrieves consumption rate data for that product category from a group of users in the same city, similar age group, and same gender within a preset time window, calculates its average value as the initial consumption rate, and gradually replaces the parameter with personalized data after the user makes their next purchase record.

[0085] The system establishes personalized item consumption curves based on initial inventory and average daily consumption rate, and calculates the estimated remaining days for each item in real time. When a user purchases the same type of product again, the system adds a new inventory record to the virtual inventory, while retaining historical consumption data for continuous optimization of consumption rate parameters.

[0086] Preferably, the estimated remaining days of use are obtained by dividing the current remaining quantity of the item by the user's average daily consumption rate of that item category.

[0087] The daily average consumption rate is determined as follows: For users with existing purchase records, the system retrieves all purchase records for this product category within a preset historical time window, extracting the quantity and corresponding purchase time for each purchase. The system calculates the total time span, starting with the first purchase time and ending with the last purchase time. The total consumption is obtained by summing the quantities of all purchase records within this time span, and then dividing the total consumption by the total time span to obtain the daily average consumption rate.

[0088] For example, a user purchased the same brand of shampoo three times in the past year: 400ml on January 1st, 400ml on April 20th, and 400ml on August 10th. The system uses the total time span of 221 days (January 1st to August 10th) as a reference, adding the total consumption of 1200ml from the three purchases to calculate the average daily consumption rate as approximately 5.43ml. Based on this, when the user purchases shampoo for the fourth time (400ml), the system divides 400ml by 5.43ml / day, estimating the remaining days to be approximately 73 days.

[0089] When a user makes their first purchase of a product category, the system uses the average daily consumption rate of similar users as the initial value, and then gradually updates it to personalized parameters as the user makes subsequent purchases.

[0090] Preferably, the system continuously monitors changes to the user's shipping address. The system stores the shipping address used by the user each time they submit an order in the address history database and compares it with the user's historical primary addresses.

[0091] When the new address does not match the city or district information in the historical address, the system determines that a migration event has occurred. The system records the migration time, origin, destination, and migration type of the migration event.

[0092] The identification of migration types is based on the following rules: if the migration time is in June or July and the place of origin is a university area and the place of destination is a workplace area, it is identified as a graduate leaving school; if the migration time is related to the time of change of the user's labor contract or social security record, it is identified as a workplace relocation; if the migration direction is from a first-tier city to a third- or fourth-tier city and the migration time is around the Spring Festival, it is identified as a return home; if the real estate transaction records of the place of origin and destination show a change from a rental address to a purchased address, it is identified as a property purchase.

[0093] The system analyzes changes in consumption characteristics before and after migration, establishing a mapping relationship between migration and consumption. For example, the system statistically analyzes the changes in purchase volume of various product categories after a large number of users migrated from northern cities to southern cities, finding a significant increase in the purchase volume of dehumidifiers and sunscreens, and based on this, establishes a migration-consumption mapping rule.

[0094] The system predicts a user's future migration probability based on historical migration patterns and external data. For example, if the system identifies a user as a college student whose current address is a university dormitory, and considering that graduation season is approaching, it predicts that the user has a high probability of migrating within the next three months and adjusts the migration status parameters in the user's virtual survival scenario in advance.

[0095] Preferably, after the system detects a user's relocation, it first obtains the city feature vector of the destination city through reverse geocoding. This feature vector includes climate, consumption, and cultural characteristics. The system matches the city feature vector of the destination city with the user's historical consumption characteristics, calculates the set of new product categories that the user may need in the new city, sorts them according to the probability of need from high to low, and generates a migration-triggered recommendation list. For example, after a user moves from Beijing to Shenzhen, the system identifies that the humidity in Shenzhen is higher than in Beijing, indicating an increased probability of the user needing dehumidifiers; the ultraviolet radiation intensity in Shenzhen is higher than in Beijing, indicating an increased probability of the user needing sunscreens; and the user's historical consumption characteristics include frequent purchases of digital product accessories. Combined with the logic of needing to rearrange the work environment after moving, the system determines that the user's probability of needing computer accessories is increased.

[0096] The system implements seasonal change-triggered recommendations by determining the climate zone type based on the latitude and longitude coordinates of the virtual character's location, combining this with the current date to determine the season, and recommending seasonal products when the season changes. For example, it recommends pollen allergy protection products in spring, sunscreen and heatstroke prevention products in summer, moisturizing and seasonal clothing in autumn, and warm clothing in winter.

[0097] The system implements recommendations triggered by changes in life stage: It analyzes the recent trends in a user's product category consumption on a weekly basis. When the purchase frequency of a particular product category changes significantly over several consecutive weeks, the system infers that the user's life stage may have changed. For example, a sudden increase in pet supply purchases from zero for several weeks suggests the user has acquired a new pet; a sudden and sustained increase in baby and maternity product purchases suggests the user has entered the parenting stage. The system then recommends suitable products based on this inferred new life stage.

[0098] Preferably, the system encodes the user's virtual life scenario features into a scenario feature vector. This vector includes the residential city code, urban area type code, community type code, consumption rate parameters for each category of goods, shopping cycle parameters, brand loyalty index, migration status identifier, and life stage identifier. The scenario feature vector is a high-dimensional sparse vector, and the value range of each component is normalized to be of the same dimension.

[0099] The system encodes users' historical interaction behavior into an interaction feature vector, which is constructed based on the number of times users purchase, browse, add to cart, favorite, and rate various products, and is also normalized.

[0100] The system concatenates the scene feature vector with the interaction feature vector to form a comprehensive user feature vector.

[0101] When calculating the similarity between users, the system adopts a weighted combination method: the cosine similarity between the two users on the scene feature vector is calculated as the scene similarity, and the cosine similarity between the two users on the interaction feature vector is calculated as the interaction similarity. Then, the scene similarity and interaction similarity are weighted and summed with preset weights to obtain the comprehensive user similarity.

[0102] Scene similarity measures the degree of similarity between two users at the level of a virtual survival scenario. Interaction similarity measures the degree of similarity between two users at the level of historical purchasing behavior. The value of scene feature weight ranges from 0.3 to 0.5, and this weight value can be adjusted according to the application scenario and dataset characteristics. When the proportion of new users in the platform is relatively high, the scene feature weight can be appropriately increased to alleviate the problem of unreliable interaction similarity caused by sparse interaction data.

[0103] The system identifies the K nearest neighbors most similar to the current user based on the overall user similarity, aggregates the ratings or interactions of these nearest neighbors with the target product, and predicts the current user's preference score for the target product.

[0104] Preferably, the system standardizes each user's virtual survival scenario data, converts it into a unified data format, and then stores it in the aggregated database. During the data entry process, the system irreversibly encrypts and de-identifies the user ID, retaining only the structured fields related to demand forecasting and analysis.

[0105] The system aggregates individual demands across time and space, generating demand forecasts at the regional and category levels. Taking the operation of an e-commerce platform in Nanshan District, Shenzhen as an example, the system scans the virtual inventory of 100,000 users in the area, extracts the estimated expiration date of each user's goods across various categories, and aggregates and statistically analyzes the data by date and geographical location.

[0106] The system generates replenishment suggestions for merchants based on aggregated demand forecasts, including the following: replenishment items, suggested replenishment quantity, suggested replenishment time, and suggested replenishment location. The system then pushes these replenishment suggestions to the corresponding merchants via their mobile applications.

[0107] The system also optimizes inventory distribution and logistics scheduling on the platform side based on demand forecasting. Specifically, this includes intelligent warehousing operations that pre-allocate goods from the central warehouse to regional forward warehouses, and precise marketing operations that configure promotional resources based on regional demand heat maps.

[0108] Preferably, for each product category, the system aggregates and statistically analyzes the estimated depletion dates of each user according to calendar dates. Specifically, the system scans the virtual inventory of all users, extracts event records for a certain product category that will be depleted on a certain date, groups and counts all records by date, and generates a time-series demand forecast curve for that product category. This curve, with date on the horizontal axis and the number of demanding users on the vertical axis, reflects the demand distribution of that product category on the time axis.

[0109] The spatial aggregation implementation method involves associating each user's geographic location information with demand forecasts. Specifically, the system aggregates data at each geographic level, merging demand forecasts for users belonging to the same geographic region to generate demand heatmaps for each level. These heatmaps use a map as their base and employ color depth or numerical labels to reflect the demand for specific product categories in each region over a specific time period.

[0110] Implementation of migration dimension aggregation: The system correlates user migration predictions with new demand at the destination. Specifically, the system batch detects migration events, identifies user groups migrating from the same origin to the same destination within the same time period, predicts and categorizes the new demand categories that this user group may generate at the destination, and generates migration-driven regional demand warnings.

[0111] Preferably, it is recommended that the replenishment quantity be determined by multiplying the forecasted demand by a safety factor. The forecasted demand is derived from the demand forecast results of a spatiotemporal aggregation model for a specific geographical region, a specific time window, and a specific product category.

[0112] The safety factor is greater than one, and its specific value is dynamically adjusted based on the following factors: the historical demand forecast error for this product category, the supply stability of this product category, and the time remaining until the peak demand period. A larger forecast error results in a higher safety factor; a more unstable supply also results in a higher safety factor; and a longer time remaining until the peak demand period also results in a higher safety factor.

[0113] The system allocates recommended replenishment quantities according to the demand forecast ratio of each sub-region in the replenishment suggestions. For example, if the predicted demand for the Science Park area is a certain value, a corresponding proportion of recommended replenishment quantity is allocated; if the predicted demand for the Nanshan Central area is a certain value, a corresponding proportion of recommended replenishment quantity is allocated; and other areas allocate the remaining recommended replenishment quantity according to their respective predicted demand ratios.

[0114] A large-scale collaborative filtering recommendation system for goods based on a hybrid recommendation algorithm. The system consists of ten modules, which communicate and collaborate with each other through a data interface.

[0115] The data acquisition module collects multi-dimensional user data in real time. This module obtains order and product data through the e-commerce platform's open API, user behavior data through a data tracking log system, and spatiotemporal data through IP location services and mobile LBS services. The collected data is cleaned, standardized, and anonymized before being written to the data warehouse.

[0116] The virtual scenario construction module reads multi-dimensional data from each user's data warehouse and constructs a virtual survival scenario for each user according to a preset construction process. The output of this module is user scenario records in the scenario database. Each record contains the user's spatial location information, virtual inventory snapshot, shopping habit parameters, and migration trajectory records.

[0117] The item consumption prediction module scans the virtual inventory records in the scenario database daily, calculates the estimated remaining days for each inventory record, and writes the calculation results back to the scenario database. When the estimated remaining days are lower than a preset threshold, the module sends a consumption warning message to the recommendation generation module.

[0118] The scene change detection module monitors the address data input by the data acquisition module in real time. When the log records show a change in the user's address at the city or district level, this module generates a migration event record. This module also performs weekly consumption structure change analysis to identify potential changes in life stages.

[0119] The recommendation generation module receives consumption warning messages and scene change events, generates consumption-triggered recommendation lists and scene change-triggered recommendation lists respectively, and sends the two recommendation results to the multi-objective optimization unit.

[0120] The collaborative filtering enhancement module reads the user's scene feature vector from the scene database and the user's interaction feature vector from the interaction database. After concatenating them to form a comprehensive feature vector, it performs collaborative filtering recommendation based on user similarity and sends the generated collaborative filtering recommendation list to the multi-objective optimization unit.

[0121] The demand aggregation module reads virtual survival scenario data from all users from the scenario database on a daily schedule, and performs time-dimensional aggregation, spatial-dimensional aggregation, and migration-dimensional aggregation in sequence to generate multi-level demand forecast results, which are then stored in the demand forecast database.

[0122] The replenishment suggestion generation module reads the aggregated demand forecast results from the demand forecast database and generates daily replenishment suggestions, promotional inventory preparation suggestions, and migration warning suggestions according to different product categories and regions. These suggestions are then pushed to the merchant's terminal through the merchant service interface.

[0123] The supply and demand matching optimization module reads demand forecast results from the demand forecast database, generates intelligent warehouse scheduling schemes and promotional resource allocation schemes, and pushes them to the platform's logistics management system and marketing management system for execution, respectively.

[0124] The feedback optimization module collects implicit and explicit user feedback on recommendation results, as well as deviation data between the merchant's adoption rate of replenishment suggestions and actual demand. It regularly performs incremental training and parameter tuning on each prediction model to achieve closed-loop continuous evolution of the system.

[0125] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0126] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0127] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0128] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0129] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0130] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0131] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0132] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0134] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A large-scale collaborative filtering recommendation method for goods based on a hybrid recommendation algorithm, characterized in that, Includes the following steps: S1: Collect multi-dimensional data of users, including shopping data, address data, behavior data and spatiotemporal data; S2: Construct a virtual character survival scenario based on the multi-dimensional data. The virtual character survival scenario includes at least: the virtual character's spatial location information, virtual inventory information, shopping habit information, and migration trajectory information. S3: Based on the virtual inventory information and the shopping habit information, calculate the estimated remaining usage days for each item; S4: When the estimated remaining usage days are lower than a preset threshold, a consumption-triggered recommendation is triggered. S5: Detect residence change events based on the migration trajectory information. When a residence change is detected, trigger scene change-triggered recommendations. S6: The features of the virtual character's survival scene are used as an enhanced dimension of the user profile and integrated into the collaborative filtering algorithm for enhanced collaborative filtering recommendations. S7: Combine the results of the consumption-triggered recommendations, the scene change-triggered recommendations, and the collaborative filtering-enhanced recommendations to generate the final recommendation list.

2. The method according to claim 1, characterized in that, The method for constructing the virtual inventory information includes: When a user purchases a product, the initial stock of that product in the virtual inventory is initialized based on the product specifications and category attributes. The user's average daily consumption rate is estimated based on the user's historical purchase intervals for this product category, and then corrected by combining it with the average consumption rate of similar users to establish a personalized item consumption curve. Based on the initial inventory and the average daily consumption rate, the estimated remaining days for each item are calculated in real time.

3. The method according to claim 1, characterized in that, The estimated remaining usage days are obtained by dividing the current remaining quantity of the item by the user's average daily consumption rate for that category. The average daily consumption rate is determined as follows: for users with existing purchase records, the average consumption per unit time is calculated based on the quantity of items purchased multiple times within a preset historical time window and the corresponding time points; for users without historical purchase records, the average consumption rate of similar users in that category is used as the initial value, and is gradually updated to personalized parameters after the user generates subsequent purchase records.

4. The method according to claim 1, characterized in that, The method for constructing the migration trajectory information includes: Continuously monitor changes in users' delivery addresses. When the new address is inconsistent with the city or district information in the historical address, record the migration event. The migration event includes migration time, place of origin, place of destination, and migration type. Analyze the changes in consumption characteristics before and after migration, and establish a mapping relationship between migration and consumption; Predict users' future migration probability based on historical migration patterns and external data.

5. The method according to claim 1, characterized in that, The scenario-change-triggered recommendations include: Migration-triggered recommendations: When a user's place of residence is detected to have changed, the city feature vector of the destination is obtained, the city feature vector is matched with the user's historical consumption characteristics, the set of new demand categories that the user may generate in the new city is calculated, and a migration-triggered recommendation list is generated. Seasonal recommendations: Recommend seasonal products based on the climate and seasonal changes in the location of the virtual character; Recommendations triggered by changes in life stage: Based on changes in the user's consumption category structure, we infer changes in life stage and recommend suitable products for the corresponding life stage.

6. The method according to claim 1, characterized in that, In the collaborative filtering enhanced recommendation, the user's virtual survival scene features are encoded as scene feature vectors, and the user's historical interaction behavior is encoded as interaction feature vectors. These are concatenated to form a comprehensive user feature vector. When calculating the similarity between users, the scene similarity of two users on the scene feature vector and the interaction similarity on the interaction feature vector are calculated separately. The scene similarity and the interaction similarity are then weighted and summed with preset weights to obtain the comprehensive user similarity. Based on the comprehensive similarity of the users, nearby users are identified, and the interaction information of the nearby users with the target product is aggregated to predict the current user's preference score for the target product.

7. The method according to claim 1, characterized in that, It also includes the following steps: After standardizing the virtual survival scenario data of each user, the data is aggregated in the spatiotemporal dimension to generate regional and category-level demand forecasts. Based on the demand forecast, replenishment suggestions are generated for merchants, including replenishment items, suggested replenishment quantity, suggested replenishment time, and suggested replenishment location; The replenishment suggestion will be pushed to the corresponding merchant terminal; Based on the aforementioned demand forecast, the platform's inventory distribution and logistics scheduling are optimized.

8. The method according to claim 7, characterized in that, The steps for aggregation in the spatiotemporal dimension include: Aggregation by time dimension: Merge and statistically analyze the estimated expiration dates of the same product or category for each user to generate a time series demand forecast curve; Aggregation by spatial dimension: Linking the geographical location information of each user with demand forecasts to generate a regional demand heat map; Aggregation by migration dimension: Correlate user migration predictions with new demand at destinations to generate migration-driven regional demand alerts.

9. The method according to claim 7, characterized in that, The recommended replenishment quantity is determined by multiplying the predicted demand by a safety factor greater than one, and its specific value is dynamically adjusted based on the historical demand forecast error and supply stability of the product category. The recommended replenishment quantity is allocated according to the demand forecast ratio of each sub-region in the replenishment recommendation.

10. A large-scale collaborative filtering recommendation system for goods based on a hybrid recommendation algorithm, applied to the method described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect multi-dimensional data from users, including shopping data, address data, behavioral data, and spatiotemporal data. The virtual scene construction module is used to construct a virtual character survival scene based on the multi-dimensional data. The virtual character survival scene includes at least the virtual character's spatial location information, virtual inventory information, shopping habit information, and migration trajectory information. The item consumption prediction module is used to calculate the estimated remaining number of days of use for each item based on the virtual inventory information and the shopping habit information; The scene change detection module is used to detect residence change events and life stage change events based on the migration trajectory information; The recommendation generation module is used to generate consumption-triggered recommendations based on the estimated remaining days of use, and to generate scenario-change-triggered recommendations based on the residence change event and life stage change event. The collaborative filtering enhancement module is used to incorporate the features of the virtual character's survival scene as an enhanced dimension of the user profile into the collaborative filtering algorithm for collaborative filtering enhancement recommendation, and to generate the final recommendation list by combining the results of consumption-triggered recommendation, scene change-triggered recommendation and collaborative filtering enhancement recommendation. The demand aggregation module is used to aggregate virtual survival scenario data of various users in the spatiotemporal dimension to generate regional and category-level demand forecasts. The replenishment suggestion generation module is used to generate replenishment suggestions for merchants based on the demand forecast. The supply and demand matching optimization module is used to optimize the inventory distribution and logistics scheduling on the platform side based on the demand forecast. The feedback optimization module is used to continuously optimize the model parameters of each module based on user and merchant feedback.

Images