Commodity recommendation method and system, storage medium and electronic equipment

Abstract

The invention discloses a commodity recommendation method and system, a storage medium and electronic equipment, and the method comprises the steps: obtaining the multi-modal data of a to-be-recommended commodity, and carrying out the normalization and feature extraction processing of the multi-modal data, so as to obtain the features of the multi-modal data; determining a data change feature and a user preference evolution feature according to the multi-modal data feature, and fusing the multi-modal data feature, the data change feature and the user preference evolution feature through a preset fusion mode to obtain a fused feature vector; determining a profit rate prediction value of the to-be-recommended commodity in the target scene according to the fusion feature vector and the multiple scene labels; and determining a target commodity combination according to the profit rate prediction value and a preset constraint condition. According to the method, by obtaining the multi-modal data and integrating and fusing the multi-modal data, the problem that product selection prediction and actual requirements are disjointed due to a single data dimension is solved, and thus the product selection accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a product recommendation method, system, storage medium, and electronic device. Background Technology

[0002] With the digital transformation of the retail industry, AI (Artificial Intelligence) models are increasingly being applied to the product selection process to improve the traditional reliance on human experience, decision-making delays, and insufficient accuracy in product selection. Among related technologies, AI techniques commonly used in the retail industry are typically applied to sales forecasting, simple recommendation algorithms, and data analysis systems for single scenarios (such as online e-commerce). However, AI-based product selection methods in these technologies often rely on historical sales data or single-market data. Due to incomplete data dimensions, the models cannot fully reflect market demand, leading to a disconnect between product selection predictions and actual needs, thereby reducing the accuracy of product selection. Summary of the Invention

[0003] This invention aims to at least partially address one of the technical problems in related technologies. To this end, the first objective of this invention is to propose a product recommendation method that, by acquiring and integrating multimodal data, solves the problem of product selection prediction being out of sync with actual needs caused by a single data dimension, thereby improving the accuracy of product selection.

[0004] A second objective of this invention is to provide a computer-readable storage medium.

[0005] The third objective of this invention is to provide an electronic device.

[0006] The fourth objective of this invention is to provide a product recommendation system.

[0007] To achieve the above objectives, a product recommendation method is proposed according to a first aspect of the present invention, comprising: acquiring multimodal data of the product to be recommended, and performing normalization and feature extraction processing on the multimodal data to obtain multimodal data features; determining data change features and user preference evolution features based on the multimodal data features, and fusing the multimodal data features, data change features, and user preference evolution features through a preset fusion method to obtain a fused feature vector; determining the profit margin prediction value of the product to be recommended in a target scenario based on the fused feature vector and multiple scenario labels; and determining a target product combination based on the profit margin prediction value and preset constraints.

[0008] According to the product recommendation method of this invention, multimodal data of the products to be recommended is acquired, and the multimodal data is normalized and feature extracted to obtain multimodal data features. Data change features and user preference evolution features are determined based on the multimodal data features. The multimodal data features, data change features, and user preference evolution features are then fused using a preset fusion method to obtain a fused feature vector. The profit margin prediction value of the products to be recommended in the target scenario is determined based on the fused feature vector and multiple scenario labels. Finally, a target product combination is determined based on the profit margin prediction value and preset constraints. Therefore, by acquiring and integrating multimodal data, the problem of product selection prediction being out of sync with actual needs caused by a single data dimension is solved, thereby improving the accuracy of product selection.

[0009] According to one embodiment of the present invention, determining the profit margin prediction value of a product to be recommended in a target scenario based on a fused feature vector and multiple scenario labels includes: inputting multiple scenario labels into a pre-trained gradient boosting tree model to obtain the selection weights for each scenario; concatenating the fused feature vector and the selection weights for each scenario to obtain concatenated features; and inputting the concatenated features into a pre-trained target prediction model to obtain the sales data prediction value of the product to be recommended within a target period in the target scenario, wherein the sales data prediction value includes the sales volume prediction value, the sales revenue prediction value, and the profit margin prediction value, and the target prediction model is trained based on an adaptive loss function corresponding to each scenario.

[0010] According to one embodiment of the present invention, the preset constraints include supply chain constraint indicators and product types. The supply chain constraint indicators include supply capacity fulfillment rate, replenishment response time, and inventory cost rate. Determining the target product mix based on the profit margin forecast and the preset constraints includes: determining the supply capacity fulfillment rate as the ratio of supplier capacity to sales forecast; determining the replenishment response time as the ratio of replenishment cycle to product turnover cycle under the target scenario; and determining the inventory cost rate as the ratio of the sum of warehousing costs and logistics costs to the sales forecast; constructing an optimization model with the profit margin forecast as the objective and the supply chain constraint indicators and product types as constraints; and solving the optimization model to obtain the target product mix.

[0011] According to one embodiment of the present invention, after determining the target product mix based on the profit margin forecast and preset constraints, the method further includes: acquiring various types of data after the target product mix is ​​implemented; and updating the gradient boosting tree model, the target prediction model, and the supply chain constraint indicators based on the difference between the various types of data and the sales data forecast.

[0012] According to one embodiment of the present invention, determining data change features and user preference evolution features based on multimodal data features includes: inputting sales data features and market environment features from the multimodal data features into a pre-trained time series feature extraction model to obtain data change features; and inputting user behavior features from the multimodal data features into a pre-trained user profile evolution model to obtain user preference evolution features.

[0013] According to an embodiment of the present invention, the method further includes: generating an evaluation report of the target product combination based on multiple types of data, wherein the evaluation report includes the actual profit margin; and updating the time-series feature extraction model, the user profile evolution model, and the preset fusion method when the ratio of the difference between the profit margin prediction value and the actual profit margin to the profit margin prediction value is greater than a preset ratio.

[0014] According to one embodiment of the present invention, multimodal data features include structured data features and unstructured data features. The process of normalizing and extracting features from the multimodal data to obtain multimodal data features includes: filling missing values ​​in the structured data of the multimodal data; removing outliers from the filled structured data to obtain target structured data; normalizing the target structured data to obtain structured data features; and extracting semantic features from the unstructured data of the multimodal data to obtain unstructured data features.

[0015] To achieve the above objectives, a computer-readable storage medium is provided according to a second aspect of the present invention, having stored thereon a computer program that, when processed by a processor, executes the product recommendation method of any of the foregoing embodiments.

[0016] According to the computer-readable storage medium of the present invention, by executing the computer program of the above-described product recommendation method, multimodal data is acquired and integrated and fused, thereby solving the problem of product selection prediction being out of touch with actual needs caused by a single data dimension, and thus improving the accuracy of product selection.

[0017] To achieve the above objectives, an electronic device is provided according to a third aspect of the present invention, including a memory, a processor, and a product recommendation program stored in the memory and executable on the processor. When the processor executes the product recommendation program, it implements the product recommendation method of any of the foregoing embodiments.

[0018] According to the electronic device of the present invention, the processor executes the program of the above-described product recommendation method, and by acquiring multimodal data and integrating and fusing the multimodal data, the problem of product selection prediction being out of touch with actual needs caused by a single data dimension is solved, thereby improving the accuracy of product selection.

[0019] To achieve the above objectives, a product recommendation system is proposed according to a fourth aspect of the present invention, comprising: a data acquisition module for acquiring multimodal data of products to be recommended; a data processing module for normalizing and extracting features from the multimodal data to obtain multimodal data features; a dynamic feature acquisition module for determining temporal features and user preference evolution features based on the multimodal data features; a multimodal fusion module for fusing the multimodal data features, temporal features, and user preference evolution features to obtain a fused feature vector; a demand prediction module for predicting the profit margin of the products to be recommended in a target scenario based on the fused feature vector and multiple scenario labels; and a decision module for determining a target product combination based on the profit margin of the products to be recommended in the target scenario and preset constraints.

[0020] According to an embodiment of the present invention, a product recommendation system acquires multimodal data of products to be recommended through a data acquisition module, normalizes and extracts features from the multimodal data through a data processing module to obtain multimodal data features, determines data change features and user preference evolution features based on the multimodal data features through a dynamic feature acquisition module, fuses the multimodal data features, data change features, and user preference evolution features through a multimodal fusion module using a preset fusion method to obtain a fused feature vector, determines the profit margin prediction value of the products to be recommended in the target scenario through a demand prediction module based on the fused feature vector and multiple scenario labels, and determines the target product combination through a decision module based on the profit margin prediction value and preset constraints. Therefore, by acquiring and integrating multimodal data, the system solves the problem of product selection prediction being out of sync with actual demand caused by a single data dimension, thereby improving the accuracy of product selection.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a product recommendation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a multimodal data acquisition and preprocessing process according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the working principle of scene adaptation according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a network structure for dynamic feature extraction and multimodal fusion according to an embodiment of the present invention; Figure 5 This is a system schematic diagram of an electronic device according to an embodiment of the present invention; Figure 6This is a schematic diagram of the structure of a product recommendation system according to an embodiment of the present invention; Figure 7 This is a schematic diagram of a product recommendation system according to another embodiment of the present invention. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0024] AI product selection solutions in related technologies are mainly divided into two categories: The first type is a time-series prediction-based product selection solution based on a single data dimension. This solution collects historical sales data of products (such as total transaction amount and sales volume) and uses time-series models such as RNN (Recurrent Neural Network) or LSTM (Long Short-Term Memory) to predict sales for a future period, determining the product selection list based on the prediction results. For example, an online e-commerce platform uses an LSTM model to analyze product sales data from the past 6 months and predicts the top 100 best-selling products for the following month as key selections. The second type is a scenario-based product selection solution based on static features. This solution targets a single scenario (such as an offline shopping mall store), extracts static attributes of products (such as price and category) and static user profiles (such as age and region), and uses traditional machine learning algorithms (such as logistic regression and random forest) to select products suitable for that scenario. For example, a certain offline chain brand selects products for its shopping mall stores in first-tier cities based solely on the price range (mid-to-high-end) and the user age group (25-40 years old).

[0025] Therefore, the two types of AI product selection solutions in related technologies have the following drawbacks: 1. Prediction bias caused by a single data dimension: Related technologies only use historical sales data or single market data, without incorporating product attributes (such as material and appearance), multimodal user behavior data (such as browsing videos and review text), and dynamic factors of the scenario (such as regional climate and emergencies). Due to the incomplete data dimensions, the model cannot fully reflect market demand, resulting in a disconnect between product selection prediction and actual demand. For example, lightweight outdoor equipment that is popular online may experience slow sales after selection because the cold climate of the offline store area was not taken into account. 2. Static feature modeling leads to strategy lag: Related technologies model product features (such as popularity) and user preferences in a static way, failing to capture dynamic evolution patterns. Since products have popularity cycles of boom and bust, and user preferences also change with policies (such as health consumption trends) and social events, static models cannot update features in real time, resulting in a lag in product selection strategy updates and missing market windows. For example, a certain health food product experienced a rapid decline in sales after being selected because it failed to capture the change in the "low sugar" consumption trend in time. 3. Poor versatility due to single-scenario adaptation: The product selection solutions of related technologies are designed only for a single retail scenario and do not take into account the differences between different scenarios. For example, online scenarios need to focus on logistics costs and repurchase rates, while offline scenarios need to focus on display space and immediate needs. Because the solutions are not adapted to the needs of multiple scenarios, the product selection solutions need to be adjusted manually when implemented (for example, when applying online product selections directly to community stores, products suitable for small space display need to be re-selected), reducing product selection efficiency. 4. Lack of supply chain coordination leads to implementation failure: Related technologies only focus on "demand forecasting" and fail to incorporate supply chain capabilities (supplier capacity, replenishment cycle, warehousing costs) into the decision-making process. Because supply chain constraints are not considered, even if a product is predicted to be a bestseller, if the supplier's capacity is insufficient or the replenishment cycle is too long, timely delivery will be impossible, resulting in the failure to achieve expected sales revenue; or, because warehousing costs are not considered, inventory backlogs may occur after product selection (for example, a predicted bestseller large furniture item experienced a 40% stockout rate in stores due to a supplier replenishment cycle as long as 30 days, ultimately resulting in sales falling short of expectations).

[0026] The following description, with reference to the accompanying drawings, outlines a product recommendation method, system, storage medium, and electronic device according to embodiments of the present invention.

[0027] Figure 1 This is a flowchart illustrating a product recommendation method according to an embodiment of the present invention. Figure 1 As shown, the product recommendation methods include: S101, acquire multimodal data of the products to be recommended, and perform normalization and feature extraction processing on the multimodal data to obtain multimodal data features.

[0028] Specifically, such as Figure 2As shown, multimodal data can be collected from retail enterprise ERP (Enterprise Resource Planning) systems, CRM (Customer Relationship Management) systems, e-commerce platform back-ends, product management systems, SCM (Supply Chain Management) systems, and third-party platforms (such as industry association data platforms). Multimodal data comprises five categories: sales data, user behavior data, product attribute data, market environment data, and supply chain data. Sales data includes historical sales volume, sales revenue, average order value, repurchase rate, and return rate. User behavior data includes browsing history (text / video / image), search keywords, review text, interaction behavior (likes / favorites / shares), and user profiles (age, region, consumption level). Product attribute data includes product categories, materials, functional parameters, price range, packaging specifications, and color characteristics (images / videos). Market environment data includes industry trend indices, competitor pricing / promotion strategies, seasonal factors, regional consumption habits, policy guidance (green consumption, healthy consumption), and unexpected events (epidemics, sporting events). Supply chain data includes supplier capacity, replenishment cycle, procurement costs, warehouse space occupancy rate, and logistics delivery timeliness. After collecting these five categories of data, preprocessing is required for subsequent use. Because these five categories of data include multiple data types, preprocessing includes normalization and feature extraction to obtain multimodal data features.

[0029] In some embodiments, multimodal data features include structured data features and unstructured data features. The process of normalizing and extracting features from multimodal data to obtain multimodal data features includes: filling missing values ​​in the structured data of the multimodal data; removing outliers from the filled structured data to obtain target structured data; normalizing the target structured data to obtain structured data features; and extracting semantic features from the unstructured data of the multimodal data to obtain unstructured data features.

[0030] Specifically, such as Figure 2As shown, multimodal data includes structured and unstructured data. Each of the five data categories mentioned above may contain both structured and unstructured data. The KNN (K-Nearest Neighbor) algorithm is used to impute missing values ​​in the structured data. Then, outliers are removed from the imputed structured data based on the 3σ principle, such as removing sales data that far exceeds the normal range. Finally, Z-score normalization is used to standardize the target structured data, bringing it to the same order of magnitude. Unstructured data includes text and image data, etc. For text data (such as review texts and search keywords), semantic features (such as core semantics like "good quality" and "slow logistics" in reviews) can be extracted using the BERT (Bidirectional Encoder Representations from Transformers) model. The BERT model is a pre-trained language model based on the Transformer architecture. For image data (product images / videos), visual features (such as color, style, and scene suitability) can be extracted using CNN (Convolutional Neural Network) and Transformer models. Search keywords can also be extracted using the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to extract core demand words (such as "low sugar" and "portable").

[0031] In the above embodiments, missing value imputation, outlier removal and standardization are performed on structured data, and semantic feature extraction is performed on unstructured data, thus realizing the preprocessing of multimodal data.

[0032] S102, determine the data change features and user preference evolution features based on the multimodal data features, and fuse the multimodal data features, data change features and user preference evolution features through a preset fusion method to obtain a fused feature vector.

[0033] Specifically, data change characteristics and user preference evolution characteristics are dynamic features. Sales data and market environment data can reflect the dynamic trends in product sales. Therefore, data change characteristics can be captured based on the corresponding data features of sales data and market environment data. These data change characteristics include cyclical changes in product sales (e.g., weekend sales are higher than weekday sales), trend changes (e.g., continuous growth in summer beverage sales), and abrupt change points (e.g., a sudden increase in sales due to promotional activities). Based on recent (e.g., the last 7 days / 30 days / 90 days) user behavior data, dynamic evolution characteristics of user preferences can be extracted, such as users shifting from "pursuing cost-effectiveness" to "emphasizing quality."

[0034] The preset fusion method can be CMAF (Cross-Modal Attention Fusion), which inputs multimodal data features, data change features, and user preference evolution features into CMAF. Through cross-modal attention weight calculation (up to 4 attention heads can be set), the heterogeneity problem of different modal data is solved, and a fusion feature vector with a unified dimension (e.g., 1024 dimensions) is output.

[0035] It should be noted that the preset fusion method is not limited to CMAF; it can also be an autoencoder method. By stacking autoencoder structures, features from different modalities (structured features, textual semantic features, visual features, etc.) are mapped to the same latent space. The model is optimized using reconstruction loss and inter-modal consistency loss to achieve the fusion of multimodal data. The advantage of this solution is its stronger adaptability to the fusion of unstructured data (such as product videos and long text reviews). In scenarios where product video features account for more than 40%, the signal-to-noise ratio of the fused features is 8% higher than the original solution, but the computational cost increases by about 15%. It is suitable for industries where product visual attributes have a significant impact on product selection (such as apparel and home furnishing retail).

[0036] S103, determine the predicted profit margin of the recommended product in the target scenario based on the fused feature vector and multiple scene labels.

[0037] Specifically, scenario tags can be determined based on actual application scenarios and user needs. For example, scenario tags can have two levels. The first-level scenario tags can include channel type, store size, geographic attributes, and consumption scenarios. The second-level scenario tags are further refined from the first-level scenario tags. For example, channel type includes online e-commerce and offline stores; store size includes community stores, standard stores, and flagship stores; geographic attributes include first-tier cities, second- and third-tier cities, and county-level markets; and consumption scenarios include daily consumption, holiday consumption, and outdoor consumption. The target scenario is the application scenario where the predicted product combination will be implemented. Based on the fusion of feature vectors and multiple scenario tags, product selection rules under different scenarios can be learned. Using the product selection rules under different scenarios, the profit margin of the recommended products under different scenarios can be predicted, thus obtaining the predicted profit margin value of the recommended products under the target scenario.

[0038] S104, Determine the target product mix based on the profit margin forecast and preset constraints.

[0039] It should be noted that the target product mix is ​​the product mix with the highest profit margin under the condition of satisfying the preset constraints. If the product mix is ​​selected solely based on profit margin forecasts, some products may have high profit margins, but they may not be available in time, resulting in the failure to achieve the expected sales revenue. Therefore, it is necessary to introduce preset constraints to ensure that the product mix can be successfully implemented, thereby achieving the expected profit.

[0040] The product recommendation method in this embodiment is not limited to recommending a target product combination based on profit margin prediction. It can also determine the target product combination based on multi-dimensional data. For example, the sales forecast value of the product to be recommended in the target scenario is determined according to the method in step S103, and the target product combination is determined based on the sales forecast value and preset constraints. The target product combination is the product combination with the highest sales.

[0041] The product recommendation method of this embodiment can also be applied to a product recommendation system, which includes a cloud server and a user terminal. The cloud server is used to execute the product recommendation method of this embodiment, and the user terminal is used to upload multimodal data to the cloud server and display the target product combination.

[0042] In the above embodiments, by acquiring and integrating multimodal data, the problem of product selection prediction being out of sync with actual needs caused by a single data dimension is solved, thereby improving the comprehensiveness and accuracy of product selection. Furthermore, the product selection accuracy can reach 86.2%, which is 23.7% higher than the product selection scheme in related technologies (accuracy rate of 62.5%). By capturing market trends and the periodic and trend changes in product sales and the evolution of user preferences, the product selection strategy update cycle is shortened, ensuring that product selection keeps up with market trends.

[0043] In some embodiments, determining the profit margin prediction value of the product to be recommended in a target scenario based on the fused feature vector and multiple scenario labels includes: inputting multiple scenario labels into a pre-trained gradient boosting tree model to obtain the selection weights for each scenario; concatenating the fused feature vector and the selection weights for each scenario to obtain concatenated features; and inputting the concatenated features into a pre-trained target prediction model to obtain the sales data prediction value of the product to be recommended within a target period in the target scenario, wherein the sales data prediction value includes the sales volume prediction value, the sales revenue prediction value, and the profit margin prediction value, and the target prediction model is trained based on the adaptive loss function corresponding to each scenario.

[0044] like Figure 3As shown, the gradient boosting tree model can be trained using the aforementioned multimodal data. Scene labels are converted into numerical features through mapping, and product selection influence features can be factors influencing product selection, such as logistics costs, repurchase rates, display space, and immediate demand. Scene label mapping and product selection influence features serve as input data, and sample labels can be defined as the true product selection weights for each scene. The gradient boosting tree model improves model accuracy by iteratively training multiple decision trees. Multiple scene labels are input into the gradient boosting tree model to obtain the product selection weights for each scene, thus deriving the product selection weight rules for each scene. For example, in online scenes, the weight of "logistics cost" is 0.2, and the weight of "repurchase rate" is 0.3; in offline community stores, the weight of "display space" is 0.25, and the weight of "immediate demand" is 0.3. Then, the fused feature vector and the product selection weights for each scene are concatenated to obtain the concatenated features. The target prediction model can be an improved lightweight Transformer model, which simplifies the encoder layer to 4 layers and sets the hidden layer dimension to 384. During training, an adaptive loss function corresponding to each scenario is introduced. For example, the prediction error penalty coefficient for offline scenarios is 1.2, and the prediction error penalty coefficient for online scenarios is 1.0. Finally, the model outputs the predicted sales volume, sales revenue, and profit margin of the recommended product within the target period under the target scenario. The target period can be the next 7 days, the next 30 days, etc. The prediction accuracy of the target prediction model can reach over 85%.

[0045] In the above embodiments, by learning the product selection weight rules for different scenarios, the product selection plan can be implemented without a lot of manual adjustments, and the implementation efficiency can be improved by more than 50%, which can be adapted to more than 80% of retail scenarios.

[0046] In some embodiments, the preset constraints include supply chain constraint indicators and product types. The supply chain constraint indicators include supply capacity fulfillment rate, replenishment response time, and inventory cost rate. Determining the target product mix based on the profit margin forecast and preset constraints includes: determining the supply capacity fulfillment rate as the ratio of supplier capacity to sales forecast; determining the replenishment response time as the ratio of replenishment cycle to product turnover cycle under the target scenario; and determining the inventory cost rate as the ratio of the sum of warehousing costs and logistics costs to the sales forecast; constructing an optimization model with the profit margin forecast as the objective and the supply chain constraint indicators and product types as constraints; and solving the optimization model to obtain the target product mix.

[0047] Specifically, the preset constraints include supply chain constraint indicators and product types. First, supply chain constraint indicators need to be constructed, including supply capacity fulfillment rate, replenishment response time, and inventory cost rate. The supply capacity fulfillment rate and inventory cost rate are calculated based on the prediction results output by the target prediction model, while the replenishment response time can be calculated based on the replenishment cycle in multimodal data and the product turnover cycle under the target scenario. Then, with the goal of maximizing the predicted profit margin, and using the supply chain constraint indicators and product types as constraints, a multi-objective optimization model is constructed. The constraints include: supply capacity fulfillment rate ≥ 85%, replenishment response time for community stores ≤ 7 days, replenishment response time for online e-commerce ≤ 15 days, inventory cost rate ≤ 30%, and the number of selected products of the same type ≤ total number of selected products * 30%. Then, a genetic algorithm (population size 100, iterations 50) can be used to solve for the target product combination, which is the optimal product combination. For example, 300 SKUs (Stock Keeping Unit) for community stores and 800 SKUs for online e-commerce.

[0048] It should be noted that particle swarm optimization (PSO) or simulated annealing can also be used to solve the problem. PSO finds the optimal solution to a multi-objective optimization problem by simulating the flight, update speed, and position of particles in space. Compared to genetic algorithms, PSO converges faster (reducing the number of iterations by 25%). In scenarios with a small number of SKUs (e.g., less than 300 SKUs in a community store), it improves efficiency by 30% and is easier to tune parameters. Simulated annealing, by simulating temperature changes during metal annealing, accepts poor solutions with a certain probability, avoiding getting trapped in local optima. In scenarios with complex product selection constraints (e.g., simultaneously considering supply chain costs, user satisfaction, and product variety), simulated annealing has a 6.2% higher probability of finding the global optimum than genetic algorithms, but its computation time increases by about 20%. It is suitable for high-end retail scenarios (e.g., luxury goods and imported goods retail) where the accuracy of product selection is extremely important.

[0049] In the above embodiments, by screening target product combinations through preset constraints, problems such as insufficient supply chain response or inventory backlog can be avoided. The success rate of product selection solutions reaches 91.5%, which is 33.2% higher than the solutions in related technologies (success rate of 58.3%).

[0050] In some embodiments, determining data change features and user preference evolution features based on multimodal data features includes: inputting sales data features and market environment features from the multimodal data features into a pre-trained time-series feature extraction model to obtain data change features; and inputting user behavior features from the multimodal data features into a pre-trained user profile evolution model to obtain user preference evolution features.

[0051] Sales data characteristics are obtained after preprocessing sales data, while market environment characteristics are obtained by predicting market environment data. Figure 4 As shown in the example, the time series feature model can be the time series attention Transformer model, which sets up multi-scale time series attention windows, such as 7 days as the daily window and 30 days as the monthly window. Then, the time series attention Transformer is used to extract time series features to obtain data change features.

[0052] User behavior characteristics are obtained after preprocessing user behavior data. User profile evolution models can be such as... Figure 4 The LSTM model shown extracts user preference evolution features from user behavior features.

[0053] Then, the multimodal data features, data change features, and user preference evolution features are input into CMAF, and CMAF outputs a fused feature vector with a unified dimension.

[0054] It's important to note that temporal feature models are not limited to the Temporal Attention Transformer model; they can also be TCN (Temporal Convolutional Network). TCN captures long-term temporal dependencies through causal convolution and dilated convolution, enabling more efficient identification of abrupt changes in sales (such as sales fluctuations caused by sudden events). Compared to the Temporal Attention Transformer model, TCN improves the accuracy of identifying sales abrupt changes by 4.1%, making it more suitable for retail scenarios with frequent market fluctuations (such as intensive holiday promotions and numerous sudden public events), and shortening the model training cycle by 20%.

[0055] The user profile evolution model is not limited to the LSTM model; it can also be a bidirectional gated recurrent unit (BRU). The BRU simplifies the cell state structure of the LSTM by using update and reset gates, reducing computational complexity (by approximately 30%) while maintaining accuracy in extracting user preference evolution features. This makes it suitable for small and medium-sized retail enterprises with limited hardware resources. In actual testing, the BRU achieved 2.3% lower accuracy in extracting dynamic user preference features compared to the LSTM model, but its inference speed was improved by 15%, meeting the needs of scenarios with high real-time requirements (such as short-cycle product selection in community group buying).

[0056] In this embodiment, the product selection strategy update cycle can be shortened from "monthly" to "weekly / daily" by using a time-series feature extraction model and a user profile evolution model, thereby avoiding missing the market window.

[0057] In some embodiments, after determining the target product mix based on the profit margin forecast and preset constraints, the method further includes: acquiring various types of data after the target product mix is ​​implemented; and updating the gradient boosting tree model, the target prediction model, and the supply chain constraint indicators based on the difference between the various types of data and the sales data forecast.

[0058] Specifically, various types of data are collected after the target product combination is implemented through Redis caching. The Redis cache can be updated hourly. These data include actual sales data, user feedback data, and supply chain status data. Actual sales data includes sales volume, sales revenue, and return rate; user feedback data includes ratings and complaints; and supply chain status data includes replenishment delay rate. Based on the difference between the various data and the predicted sales data, the FTRL-Proximal (Follow-The-Regularized-Leader Proximal) online learning algorithm is used to update the gradient boosting tree model, the target prediction model, and the supply chain constraint indicators. The update frequency can be daily to adjust the product selection weights and constraint indicators, thereby updating the target product combination.

[0059] In some embodiments, the method further includes: generating an evaluation report of the target product combination based on multiple types of data, wherein the evaluation report includes the actual profit margin; and updating the time-series feature extraction model, the user profile evolution model, and the preset fusion method when the ratio of the difference between the profit margin prediction value and the actual profit margin to the profit margin prediction value is greater than a preset ratio.

[0060] In other words, after collecting various types of data, an evaluation report on the target product combination can be generated according to a preset period (e.g., 7 days). The core indicators of the evaluation report include actual profit margin, inventory turnover rate, and user satisfaction. When the ratio of the difference between the predicted profit margin and the actual profit margin to the predicted profit margin is greater than a preset ratio, i.e., the actual profit margin is lower than the ratio of the predicted profit margin to the preset ratio, the time-series feature extraction model, the user profile evolution model, and the preset fusion method are updated to form a continuously iterative closed loop.

[0061] It should be noted that after determining the target product combination based on the sales forecast and preset constraints, the time-series feature extraction model, user profile evolution model, and preset fusion method can be updated if the ratio of the difference between the sales forecast and the actual sales to the sales forecast is greater than a preset ratio.

[0062] In the above embodiments, the product selection strategy can be optimized in real time through a closed-loop iteration mechanism, reducing the frequency of manual intervention. For example, the inventory turnover rate increases from 10.2 times / year in related technologies to 12.8 times / year, an increase of 25.5%, significantly reducing the capital occupation cost caused by inventory backlog and reducing sales losses due to stockouts. By replacing manual screening and adjustment of the product selection list with scenario adaptation and automated product selection optimization, labor costs are reduced by 62%. Assuming a sales company has 50 stores, the related technologies require a team of 3-5 people dedicated to product selection adjustment. After adopting the product recommendation method in this embodiment, only one person is needed for parameter monitoring, significantly reducing labor costs. Through the superposition of multi-dimensional optimization measures, the profit margin of retail stores increases from 24.1% in related technologies to 28.7%, an increase of 4.6 percentage points. Furthermore, as the model iterates and optimizes, the matching degree between the product selection strategy and market demand and supply chain capabilities continues to improve, and the profit margin can maintain a growth space of 15%-20%, bringing continuous economic benefits to the enterprise.

[0063] It should be noted that the product recommendation method in this embodiment can be further optimized based on specific application scenarios: Taking a multilingual scenario as an example, the above embodiments mainly target user behavior data in a Chinese environment (such as Chinese review text and Chinese search keywords), which can be expanded into a "multilingual adaptive product selection solution" to adapt to multilingual retail scenarios such as cross-border e-commerce. Specifically, during data preprocessing, multilingual pre-trained models (such as mBERT and XLM-RoBERTa) are used to replace the original single-language BERT model to achieve semantic extraction of text data such as English, Japanese, and Southeast Asian languages; "language region labels" (such as English region, Japanese region, and Spanish region) are added to the scenario labeling system to learn the differences in user preferences in different language regions (such as English-speaking users paying more attention to the environmental attributes of products, while Japanese-speaking users paying more attention to the detailed design of products). Through the above improvements, the accuracy of product selection demand prediction for cross-border e-commerce platforms can be maintained above 82%, meeting the product selection needs of multilingual markets and helping enterprises expand their overseas retail business.

[0064] Taking live-streaming e-commerce as an example, to address the characteristics of live-streaming e-commerce—"real-time interaction and rapid sales fluctuations"—a "live-streaming interaction data interface" is added during multimodal data collection to collect real-time data on user comments, likes, orders, and returns in the live stream (updating frequency increased to 10 seconds / time). A "real-time incremental learning algorithm" (such as incremental training based on online gradient descent) is introduced to shorten the model update cycle from once a day to once every 10 minutes. A "real-time sales fluctuation warning" is added, automatically triggering a supply chain replenishment reminder when the sales growth rate of a product in the live stream exceeds a preset threshold (e.g., a 20% increase within 5 minutes). Through these improvements, the product stockout rate in live-streaming e-commerce scenarios can be reduced from 18% in related technologies to 7%, while also reducing inventory backlog caused by overstocking. The profit margin related to product selection in live-streaming sessions can be increased by 5-8 percentage points, adapting to the high-frequency, real-time product selection needs of live-streaming e-commerce.

[0065] Taking the "pre-sale + self-pickup" scenario of community group buying as an example, considering the characteristics of community group buying such as "procurement based on sales and limited self-pickup point capacity," the following improvements are made: "self-pickup point capacity data" (such as the storage space of each self-pickup point and the average daily number of self-pickups) and "pre-sale cycle data" (such as 3-day pre-sale and 7-day pre-sale) are added to the supply chain data collection stage; "self-pickup point scale tags" (small community points, large community points) and "pre-sale cycle tags" are added to the scenario tag system; and "self-pickup point capacity utilization rate" (product volume / available space at the self-pickup point) and "pre-sale conversion rate constraint" (predicted pre-sale conversion rate ≥ 60%) are added to the supply chain constraint assessment indicators. Through these improvements, the out-of-stock rate of pre-sale products on the community group buying platform can be reduced to below 5%, the inventory backlog rate at self-pickup points can be reduced from 12% to 3%, while improving the user self-pickup experience (avoiding temporary storage issues due to insufficient self-pickup point capacity), and increasing the pre-sale conversion rate by 8-10 percentage points, thus meeting the "light inventory, high turnover" product selection needs of community group buying.

[0066] In summary, the product recommendation method according to embodiments of the present invention acquires multimodal data of the products to be recommended, performs normalization and feature extraction processing on the multimodal data to obtain multimodal data features, determines data change features and user preference evolution features based on the multimodal data features, and fuses the multimodal data features, data change features, and user preference evolution features through a preset fusion method to obtain a fused feature vector. Based on the fused feature vector and multiple scenario labels, the predicted profit margin of the products to be recommended in the target scenario is determined, and the target product combination is determined based on the predicted profit margin and preset constraints. Therefore, by acquiring and integrating multimodal data, the problem of product selection prediction being out of sync with actual needs caused by a single data dimension is solved, thereby improving the accuracy of product selection.

[0067] Corresponding to the above embodiments, embodiments of the present invention also propose a computer-readable storage medium storing a computer program thereon, which, when processed by a processor, executes the product recommendation method of any of the foregoing embodiments.

[0068] According to the computer-readable storage medium of the present invention, by executing the computer program of the above-described product recommendation method, multimodal data is acquired and integrated and fused, thereby solving the problem of product selection prediction being out of touch with actual needs caused by a single data dimension, and thus improving the accuracy of product selection.

[0069] Corresponding to the above embodiments, embodiments of the present invention also propose an electronic device. For example... Figure 5 As shown, the electronic device 100 includes a memory 110, a processor 120, and a product recommendation program stored on the memory 110 and executable on the processor 120. When the processor 120 executes the product recommendation program, it implements the product recommendation method of any of the foregoing embodiments.

[0070] According to the electronic device of the present invention, the processor executes the program of the above-described product recommendation method, and by acquiring multimodal data and integrating and fusing the multimodal data, the problem of product selection prediction being out of touch with actual needs caused by a single data dimension is solved, thereby improving the accuracy of product selection.

[0071] Corresponding to the above embodiments, embodiments of the present invention also propose a product recommendation system. For example... Figure 6 As shown, the product recommendation system includes: a data acquisition module 10, a data processing module 20, a dynamic feature acquisition module 30, a multimodal fusion module 40, a demand prediction module 50, and a decision-making module 60.

[0072] The system comprises the following modules: a data acquisition module 10 for acquiring multimodal data of the products to be recommended; a data processing module 20 for normalizing and extracting features from the multimodal data to obtain multimodal data features; a dynamic feature acquisition module 30 for determining time-series features and user preference evolution features based on the multimodal data features; a multimodal fusion module 40 for fusing the multimodal data features, time-series features, and user preference evolution features to obtain a fused feature vector; a demand prediction module 50 for predicting the profit margin of the products to be recommended in the target scenario based on the fused feature vector and multiple scenario labels; and a decision module 60 for determining the target product combination based on the profit margin of the products to be recommended in the target scenario and preset constraints.

[0073] In some embodiments, such as Figure 7As shown, the product recommendation system also includes a scenario adaptation module 70. The scenario adaptation module 70 is used to input multiple scenario labels into a pre-trained gradient boosting tree model to obtain the product selection weights for each scenario; it concatenates the fused feature vector and the product selection weights for each scenario to obtain concatenated features; the demand prediction module 50 is also used to input the concatenated features into a pre-trained target prediction model to obtain the sales data prediction values ​​of the product to be recommended within the target period in the target scenario. The sales data prediction values ​​include sales volume prediction values, sales revenue prediction values, and profit margin prediction values. The target prediction model is trained based on the adaptive loss function corresponding to each scenario.

[0074] In some embodiments, the preset constraints include supply chain constraint indicators and product types. The supply chain constraint indicators include supply capacity fulfillment rate, replenishment response time, and inventory cost rate. The decision module 60 is further configured to: determine that the supply capacity fulfillment rate is the ratio of the supplier's production capacity to the sales forecast, determine that the replenishment response time is the ratio of the replenishment cycle to the product turnover cycle under the target scenario, and determine that the inventory cost rate is the ratio of the sum of warehousing costs and logistics costs to the sales forecast; construct an optimization model with the goal of maximizing the profit margin forecast and with the supply chain constraint indicators and product types as constraints; and solve the optimization model to obtain the target product combination.

[0075] In some embodiments, such as Figure 7 As shown, the product recommendation system also includes an update and feedback module 80. The update and feedback module 80 is used to obtain various types of data after the target product combination is determined based on the profit margin prediction value and preset constraints; and to update the gradient boosting tree model, the target prediction model and the supply chain constraint indicators based on the difference between the various types of data and the sales data prediction value.

[0076] In some embodiments, the dynamic feature acquisition module 30 is further configured to: input sales data features and market environment features from multimodal data features into a pre-trained time-series feature extraction model to obtain data change features; and input user behavior features from multimodal data features into a pre-trained user profile evolution model to obtain user preference evolution features.

[0077] In some embodiments, the update and feedback module 80 is further configured to: generate an evaluation report of the target product combination based on multiple types of data, wherein the evaluation report includes the actual profit margin; and update the time-series feature extraction model, the user profile evolution model, and the preset fusion method when the ratio of the difference between the profit margin prediction value and the actual profit margin to the profit margin prediction value is greater than a preset ratio.

[0078] In some embodiments, the multimodal data features include structured data features and unstructured data features. The data processing module 20 is further configured to: fill in missing values ​​in the structured data of the multimodal data; remove outliers from the filled structured data to obtain target structured data; normalize the target structured data to obtain structured data features; and extract semantic features from the unstructured data of the multimodal data to obtain unstructured data features.

[0079] It should be noted that the specific implementation of the product recommendation system in this embodiment of the invention corresponds one-to-one with the specific implementation of the product recommendation method in the foregoing embodiments of the invention, and will not be repeated here.

[0080] According to an embodiment of the present invention, a product recommendation system acquires multimodal data of products to be recommended through a data acquisition module, normalizes and extracts features from the multimodal data through a data processing module to obtain multimodal data features, determines data change features and user preference evolution features based on the multimodal data features through a dynamic feature acquisition module, fuses the multimodal data features, data change features, and user preference evolution features through a multimodal fusion module using a preset fusion method to obtain a fused feature vector, determines the profit margin prediction value of the products to be recommended in the target scenario through a demand prediction module based on the fused feature vector and multiple scenario labels, and determines the target product combination through a decision module based on the profit margin prediction value and preset constraints. Therefore, by acquiring and integrating multimodal data, the system solves the problem of product selection prediction being out of sync with actual demand caused by a single data dimension, thereby improving the accuracy of product selection.

[0081] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0082] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0083] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0084] Furthermore, the terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying relative importance, or implicitly specifying the number of technical features indicated in this embodiment. Therefore, features defined with terms such as "first" and "second" in the embodiments of this invention can explicitly or implicitly indicate that the embodiment includes at least one of those features. In the description of this invention, the word "multiple" means at least two or more, such as two, three, four, etc., unless otherwise explicitly specified in the embodiments.

[0085] In this invention, unless otherwise explicitly specified or limited in the embodiments, the terms "installation," "connection," "joining," and "fixing" appearing in the embodiments should be interpreted broadly. For example, a connection can be a fixed connection, a detachable connection, or an integral part; it can also be a mechanical connection, an electrical connection, etc. Of course, it can also be a direct connection, or an indirect connection through an intermediate medium, or it can be the internal communication of two components, or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific implementation.

[0086] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A product recommendation method, characterized in that, include: Acquire multimodal data of the products to be recommended, and perform normalization and feature extraction processing on the multimodal data to obtain multimodal data features; Based on the multimodal data features, data change features and user preference evolution features are determined, and the multimodal data features, data change features, and user preference evolution features are fused using a preset fusion method to obtain a fused feature vector; The profit margin prediction value of the recommended product in the target scenario is determined based on the fused feature vector and multiple scene labels. The target product mix is ​​determined based on the predicted profit margin and preset constraints.

2. The method according to claim 1, characterized in that, Based on the fused feature vector and multiple scene labels, the predicted profit margin of the product to be recommended in the target scene is determined, including: The various scene labels are input into a pre-trained gradient boosting tree model to obtain the product selection weights for each scene; The fused feature vector and the product selection weights for each scenario are concatenated to obtain the concatenated features; The spliced ​​features are input into a pre-trained target prediction model to obtain the sales data prediction value of the product to be recommended within the target period under the target scenario. The sales data prediction value includes the sales volume prediction value, the sales revenue prediction value and the profit margin prediction value. The target prediction model is trained based on the adaptive loss function corresponding to each scenario.

3. The method according to claim 2, characterized in that, The preset constraints include supply chain constraint indicators and product categories. The supply chain constraint indicators include supply capacity fulfillment rate, replenishment response time, and inventory cost rate. Determining the target product mix based on the profit margin forecast and preset constraints includes: The supply capacity fulfillment rate is determined to be the ratio of the supplier's production capacity to the sales forecast, the replenishment response time is determined to be the ratio of the replenishment cycle to the product turnover cycle under the target scenario, and the inventory cost rate is determined to be the ratio of the sum of warehousing costs and logistics costs to the sales forecast. Construct an optimization model with the objective of maximizing the predicted profit margin, and with the supply chain constraint indicators and the product types as constraints; The optimization model is solved to obtain the target product combination.

4. The method according to claim 3, characterized in that, After determining the target product mix based on the profit margin forecast and preset constraints, the method further includes: Obtain various types of data after the target product combination is implemented; The gradient boosting tree model, the target prediction model, and the supply chain constraint indicators are updated based on the differences between the various types of data and the predicted sales data.

5. The method according to claim 4, characterized in that, Based on the multimodal data features, data change characteristics and user preference evolution characteristics are determined, including: The sales data features and market environment features from the multimodal data features are input into a pre-trained time-series feature extraction model to obtain the data change features; The user behavior features in the multimodal data features are input into a pre-trained user profile evolution model to obtain the user preference evolution features.

6. The method according to claim 5, characterized in that, The method further includes: An evaluation report for the target product combination is generated based on the various types of data, wherein the evaluation report includes the actual profit margin; If the ratio of the difference between the predicted profit margin and the actual profit margin to the predicted profit margin is greater than a preset ratio, the time-series feature extraction model, the user profile evolution model, and the preset fusion method are updated.

7. The method according to any one of claims 1-6, characterized in that, The multimodal data features include structured data features and unstructured data features. The multimodal data undergoes normalization and feature extraction processing to obtain the multimodal data features, including: Missing values ​​in the structured data of the multimodal data are filled in; Outliers are removed from the filled structured data to obtain the target structured data; The target structured data is normalized to obtain the structured data features; Semantic features are extracted from the unstructured data in the multimodal data to obtain the features of the unstructured data.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when processed by a processor, executes the product recommendation method as described in any one of claims 1-7.

9. An electronic device, characterized in that, The device includes a memory, a processor, and a product recommendation program stored in the memory and executable on the processor. When the processor executes the product recommendation program, it implements the product recommendation method according to any one of claims 1-7.

10. A product recommendation system, characterized in that, include: The data acquisition module is used to acquire multimodal data of the products to be recommended; The data processing module is used to normalize and extract features from the multimodal data to obtain multimodal data features. The dynamic feature acquisition module is used to determine time-series features and user preference evolution features based on the multimodal data features; A multimodal fusion module is used to fuse the multimodal data features, the temporal features, and the user preference evolution features to obtain a fused feature vector; The demand prediction module is used to predict the profit margin of the recommended product in the target scenario based on the fused feature vector and multiple scenario labels. The decision module is used to determine the target product combination based on the profit margin of the products to be recommended in the target scenario and preset constraints.

Images