Artificial intelligence recommendation method and device based on commodity life cycle
By collecting and standardizing data from multiple channels, and combining fusion neural networks and multi-task learning models, the product lifecycle stages are accurately identified, and personalized recommendation schemes and operational strategies are generated. This solves the problems of single data dimensions and lagging stage determination in traditional methods, and improves recommendation accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANENG ENERGY & COMM HLDG CO LTD
- Filing Date
- 2026-04-24
- Publication Date
- 2026-06-12
Smart Images
Figure CN122199115A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to an artificial intelligence recommendation method and apparatus based on the product lifecycle. Background Technology
[0002] Product recommendation, a core driver of e-commerce and retail, is widely used in personalized marketing scenarios. With the development of deep learning technology, intelligent recommendation systems have been constructed through the collaborative work of data collection, feature mining, and model inference. Specifically, this intelligent recommendation system covers the entire process from user behavior analysis to product matching, including key stages such as transaction data integration, static feature extraction, and single-task model prediction, aiming to improve the efficiency of supply and demand matching.
[0003] However, existing recommendation methods directly use single transaction data from within the platform, resulting in a lack of awareness of the product lifecycle status. This makes the recommendation model unable to adapt to the operational needs at different stages, leading to delayed recommendation timing and insufficient strategy targeting, thus reducing the accuracy of product recommendations. Summary of the Invention
[0004] The present invention proposes an artificial intelligence recommendation method and apparatus based on the product lifecycle to solve one of the technical problems in related technologies.
[0005] To achieve the above objectives, this invention proposes an artificial intelligence recommendation method based on the product lifecycle, comprising:
[0006] Product data is collected through multiple data collection channels, and the product data is standardized to obtain a standardized multi-source fusion data set, which includes product end-to-end data, user demand data, and scenario time-series data. Product lifecycle features are obtained by extracting features from the entire product lifecycle data, and spatial and temporal correlation analysis is performed on the product lifecycle features using a fusion neural network model to obtain a product lifecycle feature vector. The product lifecycle feature vector is input into the lifecycle stage identification model to obtain the current lifecycle stage of the product; Feature extraction is performed on the user demand data and the scenario time series data to obtain user demand features and scenario time series features. The product life cycle feature vector, the user demand features and the scenario time series features are then subjected to dimension alignment and tensor fusion processing to obtain a multi-dimensional feature matrix. The multi-dimensional feature matrix and the current lifecycle stage of the product are input into a multi-task learning model to generate personalized product recommendation schemes and operational strategies adapted to the current lifecycle stage.
[0007] The AI recommendation method based on the product lifecycle in this invention may also have the following additional technical features: In one embodiment of the present invention, the standardization processing of the collected commodity data to obtain a standardized multi-source fusion data set includes: The missing values in the product collection data are filled in to obtain the filled data; The filled data is subjected to anomaly detection and removal to obtain data after removing outliers; The data after removing outliers is subjected to unitization processing to obtain normalized data; The sensitive information in the normalized data is encrypted to obtain the desensitized data; The data belonging to the same dimension or the same product entity in the de-identified data are integrated to obtain a standardized multi-source fusion data set.
[0008] In one embodiment of the present invention, the fused neural network model includes a first CNN network, an LSTM network, and an attention mechanism; the step of extracting features from the full-link data of the goods to obtain product lifecycle features, and using the fused neural network model to perform spatial and temporal correlation analysis on the product lifecycle features to obtain a product lifecycle feature vector, includes: Product lifecycle features are obtained by extracting features from the entire product chain data, wherein the product lifecycle features include transaction features, inventory features, operational features and environmental features; The transaction features, inventory features, operational features, and environmental features are dimensionally aligned and concatenated to obtain an initial feature vector; The initial feature vector is input into the first CNN network to extract the spatial correlation between features, thus obtaining the first feature vector; The first feature vector is input into the LSTM network to extract the temporal correlation of features, thus obtaining the second feature vector; The first feature vector and the second feature vector are weighted by the attention mechanism to obtain the product lifecycle feature vector.
[0009] In one embodiment of the present invention, the step of performing dimensional alignment and tensor fusion processing on the product lifecycle feature vector, the user demand features, and the scenario temporal features to obtain a multi-dimensional feature matrix includes: The user demand keywords are extracted from the user demand features using the TF-IDF algorithm or collaborative filtering algorithm, and the user demand keywords are transformed into user demand feature vectors. The temporal trend features in the temporal features of the scene are extracted by time series analysis algorithm, and the temporal trend features are transformed into a scene temporal feature vector; The Min-Max normalization method is used to transform the product lifecycle feature vector, the user demand feature vector, and the scenario time series feature vector into a zero-to-one interval and align the sample number with the time dimension to obtain the dimension-aligned product lifecycle feature vector, user demand feature vector, and scenario time series feature vector. The product lifecycle feature vector, user demand feature vector, and scenario time-series feature vector, which are aligned according to the dimensions, are fused to construct a multi-dimensional feature matrix.
[0010] In one embodiment of the present invention, the multi-task learning model includes a shared layer, a task-specific layer, and an output layer; the step of inputting the multi-dimensional feature matrix and the current lifecycle stage of the product into the multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current lifecycle stage includes: The shared layer uses a Transformer encoder to extract deep correlation features from the multi-dimensional feature matrix; The task-specific layer processes the deep correlation features and the current lifecycle stage of the product to obtain product preference scores and operational strategy types and parameters; The output layer normalizes the product preference scores and the operational strategy types and parameters to generate personalized product recommendation schemes and operational strategies adapted to the current lifecycle stage.
[0011] In one embodiment of the present invention, the task-specific layer includes a recommendation task subnetwork and an operation strategy task subnetwork; the task-specific layer processes the deep correlation features and the current lifecycle stage of the product to obtain product preference scores and operation strategy types and parameters, including: The deep association features are input into the fully connected neural network in the recommendation task subnetwork to obtain the product preference score; The product preference score and the current lifecycle stage of the product are input into the second CNN network in the operation strategy task sub-network to obtain the operation strategy type and parameters that are adapted to the current lifecycle stage of the product.
[0012] To achieve the above objectives, another aspect of the present invention proposes an artificial intelligence recommendation device based on the product lifecycle, comprising: The data acquisition module is used to acquire product data through multiple data acquisition channels, and to perform standardized processing on the product data to obtain a standardized multi-source fusion data set, wherein the multi-source fusion data set includes product full-link data, user demand data and scenario time-series data; The feature extraction module is used to extract features from the full-chain data of the product to obtain product lifecycle features, and to use a fusion neural network model to perform spatial and temporal correlation analysis on the product lifecycle features to obtain a product lifecycle feature vector. The first data processing module is used to input the product lifecycle feature vector into the lifecycle stage identification model to obtain the current lifecycle stage of the product. The second data processing module is used to extract features from the user demand data and the scenario time series data to obtain user demand features and scenario time series features, and to perform dimension alignment and tensor fusion processing on the product life cycle feature vector, the user demand features and the scenario time series features to obtain a multi-dimensional feature matrix. The generation module is used to input the multi-dimensional feature matrix and the current life cycle stage of the product into the multi-task learning model to generate personalized product recommendation schemes and operational strategies adapted to the current life cycle stage.
[0013] Another object of the present invention is to provide an electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in any one of the preceding aspects.
[0014] Another object of the present invention is to provide a computer storage medium storing computer-executable instructions; said computer-executable instructions, when executed by a processor, cause the computer to perform any of the methods described in any of the preceding aspects.
[0015] The AI-based recommendation method and apparatus based on the product lifecycle of this invention acquires product data through multiple data collection channels, standardizes the product data to obtain a standardized multi-source fusion data set, which includes product end-to-end data, user demand data, and scenario time-series data. Product lifecycle features are extracted from the product end-to-end data, and spatial and temporal correlation analyses are performed on these features using a fusion neural network model to obtain a product lifecycle feature vector. This product lifecycle feature vector is then input into a lifecycle stage identification model to determine the current lifecycle stage of the product. User demand features and scenario time-series features are extracted from the user demand data and scenario time-series data, and the product lifecycle feature vector, user demand features, and scenario time-series features are dimensionally aligned and fused using tensors to obtain a multi-dimensional feature matrix. Finally, the multi-dimensional feature matrix and the current lifecycle stage of the product are input into a multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current lifecycle stage. This invention accurately identifies product lifecycle stages by collecting product data and uses a multi-task learning model to collaboratively generate personalized product recommendation schemes and operational strategies for the current lifecycle stage. This solves the problems of single data dimensions and lagging stage determination in traditional methods, thereby improving the accuracy of product recommendations.
[0016] 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
[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating an AI-powered recommendation method based on the product lifecycle according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an AI recommendation device based on the product lifecycle according to an embodiment of the present invention. Detailed Implementation
[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.
[0020] The following description, with reference to the accompanying drawings, describes an AI-based recommendation method and apparatus based on the product lifecycle according to embodiments of the present invention.
[0021] Figure 1 This is a flowchart illustrating an AI recommendation method based on the product lifecycle according to an embodiment of the present invention. Figure 1 As shown, the method may include the following steps: Step 101: Obtain product data through multiple data collection channels, and perform standardized processing on the product data to obtain a standardized multi-source fusion data set. The multi-source fusion data set includes product end-to-end data, user demand data, and scenario time-series data.
[0022] In one embodiment of the present invention, product data is acquired through multiple data collection channels to solve the technical problem in traditional recommendation technology where the data source is singular and the dimensions are missing, resulting in an inability to fully reflect the product lifecycle status and the user's real needs.
[0023] In one embodiment of the present invention, a mechanism for the collection and standardized processing of multi-source heterogeneous data is constructed. By deploying multiple data collection channels in parallel, the mechanism comprehensively acquires product collection data covering the entire product chain, user needs, and scenario time series. The product collection data is then subjected to unified standardized processing to eliminate data noise, magnitude differences, and privacy risks, resulting in a standardized multi-source fusion data set.
[0024] In one embodiment of the present invention, the aforementioned multiple data acquisition channels logically encompass protocol interfacing with external system interfaces, direct access to internal platform databases, and signal acquisition from physical sensing devices, to ensure the real-time nature and completeness of data acquisition.
[0025] Specifically, in one embodiment of the present invention, the input source is an external data source, namely an e-commerce platform, a supply chain management system, and a third-party data service provider, through API interface connection. The processing actions include establishing an HTTPS protocol transmission channel, using a token identity authentication mechanism for security verification, and then parsing and extracting basic product information, product transaction data, basic user information, user behavior data, and industry benchmark data, and outputting a first type of data stream that has been standardized in format.
[0026] Furthermore, in one embodiment of the present invention, through platform database integration, the input source is structured and unstructured data from the enterprise's internal platform database. The processing actions include establishing a read-only connection through JDBC or ODBC protocols, executing SQL statements to accurately query structured transaction details and user profiles, and simultaneously applying text parsing and image feature extraction algorithms to unstructured user review text and product images to convert them into structured formats, thereby obtaining product inventory data, order detail data, user historical consumption data, and product operation log data, and outputting a second type of data stream.
[0027] Furthermore, in one embodiment of the present invention, data is collected through IoT devices, with input sources including deployed RFID readers, smart shelves, temperature and humidity sensors, and customer flow statistics devices. The processing actions include preprocessing the collected commodity warehousing data, offline store data, commodity environment data, and commodity logistics data through an edge computing module by denoising, filtering, and normalizing, and transmitting the processed data to the central node via the MQTT protocol to output a third type of data stream.
[0028] Furthermore, in one embodiment of the present invention, the above three types of data streams are aggregated into commodity collection data for unified cleaning and standardization processing, so as to output a standardized multi-source fusion data set as a unified input basis for subsequent feature extraction and model construction.
[0029] Specifically, in one embodiment of the present invention, the method for standardizing commodity collection data to obtain a standardized multi-source fusion data set may include the following steps: Step 1011: Fill in the missing values in the product collection data to obtain the filled data; Step 1012: Perform anomaly detection and removal on the filled data to obtain the data after removing outliers; Step 1013: Perform dimensional unification processing on the data after removing outliers to obtain normalized data; Step 1014: Encrypt sensitive information in the normalized data to obtain desensitized data; Step 1015: Integrate the data belonging to the same dimension or the same product entity from the de-identified data to obtain a standardized multi-source fusion data set.
[0030] In one embodiment of the present invention, missing values in the product collection data can be filled by means filling, median filling or machine learning prediction filling methods according to the data type to obtain the filled data.
[0031] In one embodiment of the present invention, the 3σ principle or box plot analysis algorithm can be used to detect and remove anomalies in the filled data to obtain data after removing outliers.
[0032] In one embodiment of the present invention, the Min-Max normalization method is used to transform the data after outlier removal into the same interval to eliminate feature weight imbalance, resulting in normalized data. Furthermore, in another embodiment of the present invention, sensitive information (such as user privacy data and enterprise sensitive data) in the normalized data is encrypted to obtain de-identified data.
[0033] In one embodiment of the present invention, data belonging to the same dimension or the same product entity in the desensitized data are integrated to obtain a standardized multi-source fusion data set, which serves as a unified input basis for subsequent feature extraction and model construction.
[0034] In one embodiment of the present invention, the data in the standardized multi-source fusion data set is classified. The full-chain data of the product may include basic product information, product transaction data, product inventory data, order details data, product operation log data, product warehousing data, product logistics data, and product environment data. The user demand data may include basic user information, user behavior data, and user historical consumption data. The scenario time series data may include industry benchmark data and offline store data.
[0035] In one embodiment of the present invention, the comprehensiveness, real-time nature and accuracy of data sources are ensured by using three parallel acquisition channels and a refined preprocessing process. This effectively solves the problem of single data dimension in traditional recommendation methods. Furthermore, the unified standardized processing eliminates the magnitude differences and noise interference of multi-source heterogeneous data, providing high-quality data support for subsequent accurate identification of product lifecycle stages and construction of multi-dimensional feature matrices.
[0036] Step 102: Extract features from the entire product lifecycle data to obtain product lifecycle features, and use a fusion neural network model to perform spatial and temporal correlation analysis on the product lifecycle features to obtain product lifecycle feature vectors.
[0037] In one embodiment of the present invention, after obtaining the full-chain data of the product through the above steps, the full-chain data of the product can be used to extract features to obtain product lifecycle features, and a fusion neural network model can be used to perform spatial correlation and temporal correlation analysis on the product lifecycle features to obtain product lifecycle feature vectors.
[0038] Specifically, in one embodiment of the present invention, the aforementioned fused neural network model may include a first CNN network, an LSTM network, and an attention mechanism. In another embodiment of the present invention, the method for extracting features from the entire product lifecycle data to obtain product lifecycle features, and then using the fused neural network model to perform spatial and temporal correlation analysis on these product lifecycle features to obtain a product lifecycle feature vector, may include the following steps: Step 1021: Extract product lifecycle features from the entire product supply chain data; Step 1022: Align and concatenate the transaction features, inventory features, operational features, and environmental features according to their dimensions to obtain the initial feature vector; Step 1023: Input the initial feature vector into the first CNN network to extract the spatial correlation between features and obtain the first feature vector; Step 1024: Input the first feature vector into the LSTM network to extract the temporal correlation of features and obtain the second feature vector; Step 1025: Weights are assigned to the first and second feature vectors using an attention mechanism to obtain the product lifecycle feature vector.
[0039] Specifically, in one embodiment of the present invention, the aforementioned product lifecycle characteristics may include transaction characteristics, inventory characteristics, operational characteristics, and environmental characteristics. Specifically, in one embodiment of the present invention, transaction characteristics can be extracted based on product transaction data, order detail data, and product operation log data. These transaction characteristics may include sales growth rate, sales revenue growth rate, repurchase rate, average order value, and conversion rate. Inventory characteristics can be extracted based on product inventory data, order detail data, product warehousing data, and product logistics data. These inventory characteristics may include inventory turnover rate, inventory backlog duration, replenishment frequency, and loss rate. Operational characteristics can be extracted based on product operation log data. These operational characteristics may include exposure, click-through rate, promotional investment ratio, evaluation rating, and negative review rate. Environmental characteristics can be extracted based on product warehousing data, product environmental data, product logistics data, product transaction data, and order detail data. These environmental characteristics may include storage duration, temperature and humidity compliance rate, and logistics damage rate.
[0040] In one embodiment of the present invention, the transaction features, inventory features, operational features and environmental features obtained through the above steps are dimensionally aligned, and an initial feature vector is formed by vector concatenation operation.
[0041] In one embodiment of the present invention, the initial feature vector is input into a first CNN network, and the convolutional kernel is used to perform sliding calculations on the feature dimension to extract the spatial correlation between different operational indicators and inventory status, and output a first feature vector containing spatial context information. The first feature vector is input into an LSTM network, and its gating mechanism is used to process time series data to capture the temporal correlation of key indicators such as sales growth rate and repurchase rate over time, and output a second feature vector containing the temporal evolution law. The first feature vector and the second feature vector are weighted through an attention mechanism, and by calculating the importance coefficient of each feature channel, the features that have a greater impact on the life cycle stage are highlighted. For example, the weight of sales growth rate is increased during the growth stage, and the weight of repurchase rate is increased during the decline stage, and the fused product life cycle feature vector is output.
[0042] Step 103: Input the product lifecycle feature vector into the lifecycle stage identification model to obtain the current lifecycle stage of the product.
[0043] In one embodiment of the present invention, after obtaining the product lifecycle feature vector through the above steps, the product lifecycle feature vector can be input into the lifecycle stage identification model to obtain the current lifecycle stage of the product.
[0044] In one embodiment of the present invention, the aforementioned life cycle stage identification model can be a classifier composed of a fully connected neural network. Specifically, in one embodiment of the present invention, the product life cycle feature vector is input into a classifier composed of a fully connected neural network. After nonlinear mapping processing, the probability distribution of the product being in the introduction, growth, maturity, or decline stage is output. The stage with the highest probability value is selected as the identification result of the current life cycle, and the confidence level of the result is calculated. When the confidence level is less than 85%, a manual verification mechanism is automatically triggered to ensure the accuracy of the judgment.
[0045] Furthermore, in one embodiment of the present invention, the changing trend of the probability distribution of the stages can also be monitored in real time. When the probability of a certain target stage is detected to be rising continuously within a preset time and the corresponding transaction or inventory characteristics reach the preset stage migration threshold, stage migration warning information is immediately generated. This information includes the product ID to be migrated, the current stage, the target migration stage, the warning reason and the warning level, and is pushed to the operator's terminal, thereby realizing a complete closed-loop processing from feature fusion, stage identification to migration warning.
[0046] Step 104: Extract features from user demand data and scenario time-series data to obtain user demand features and scenario time-series features. Then, perform dimension alignment and tensor fusion processing on the product lifecycle feature vector, user demand features and scenario time-series features to obtain a multi-dimensional feature matrix.
[0047] In one embodiment of the present invention, after obtaining user demand data and scenario time-series data through the above steps, feature extraction can be performed on the user demand data and scenario time-series data to obtain user demand features and scenario time-series features. The product lifecycle feature vector, user demand features and scenario time-series features are then subjected to dimension alignment and tensor fusion processing to obtain a multi-dimensional feature matrix.
[0048] In one embodiment of the present invention, user demand features are obtained by feature extraction from user demand data. In this embodiment, the user demand features may include static features and dynamic features. Static features may include age, gender, consumption level, region, and preferred product categories, while dynamic features may include recent browsing history, shopping cart additions, purchase frequency, review keywords, and demand change trends.
[0049] In one embodiment of the present invention, feature mining is performed on the scene time series data to obtain scene time series features. The scene time series features may include time features, scene features and time series correlation features. Time features include seasons, holidays and time periods. Scene features include online or offline, warehousing or stores or logistics. Time series correlation features include the time series changes of products in different scenarios, the changes of user needs in different time periods, and the correlation between the product category life cycle and the scenario.
[0050] Furthermore, in one embodiment of the present invention, the method for performing dimensional alignment and tensor fusion processing on the product lifecycle feature vector, user demand features, and scenario temporal features to obtain a multi-dimensional feature matrix may include the following steps: Step 1041: Extract user demand keywords from user demand features using the TF-IDF algorithm or collaborative filtering algorithm, and transform user demand keywords into user demand feature vectors; Step 1042: Extract the temporal trend features from the scene temporal features using a time series analysis algorithm, and transform the temporal trend features into a scene temporal feature vector; Step 1043: The Min-Max normalization method is used to transform the product lifecycle feature vector, user demand feature vector and scenario time series feature vector into the zero-to-one interval and the sample number and time dimension are aligned to obtain the dimension-aligned product lifecycle feature vector, user demand feature vector and scenario time series feature vector. Step 1044: Using tensor fusion, the dimension-aligned product lifecycle feature vector, user demand feature vector, and scenario time-series feature vector are fused to construct a multi-dimensional feature matrix.
[0051] In one embodiment of the present invention, the time series analysis algorithm described above may be ARIMA.
[0052] In one embodiment of the present invention, after obtaining the product lifecycle feature vector, user demand feature vector, and scenario time-series feature vector through the above steps, the Min-Max normalization method can be used to transform all values in the product lifecycle feature vector, user demand feature vector, and scenario time-series feature vector into the range of zero to one, so as to eliminate the influence of different feature magnitudes. At the same time, the three types of features are dimensionally aligned to ensure that the number of samples is consistent and the time dimension is synchronized. For example, the time-series features are aligned by day or by week to obtain the dimensionally aligned product lifecycle feature vector, user demand feature vector, and scenario time-series feature vector.
[0053] Furthermore, in one embodiment of the present invention, assuming that the dimension of the dimension-aligned product lifecycle feature vector X is m, the dimension of the user demand feature vector Y is n, and the dimension of the scenario temporal feature vector Z is p, the three are constructed into a three-dimensional tensor with dimensions m×n×p using a tensor fusion method, and the three-dimensional tensor is transformed into a two-dimensional feature matrix with dimensions (m×n) rows and p columns using the PARAFAC decomposition algorithm to obtain a multi-dimensional feature matrix.
[0054] In one embodiment of the present invention, this implementation method refines the mining of static and dynamic user demand features and multi-dimensional scene time-series features, and combines Min-Max normalization and PARAFAC decomposition algorithms to achieve a quantitative expression of the deep relationship between products, users and scenarios. This effectively solves the problem of single dimension in traditional feature fusion, provides high-purity and strongly correlated input data for multi-task learning models, and significantly improves the accuracy of recommendation schemes and the adaptability of operational strategies.
[0055] Step 105: Input the multi-dimensional feature matrix and the current life cycle stage of the product into the multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current life cycle stage.
[0056] In one embodiment of the present invention, after obtaining the multi-dimensional feature matrix and the current life cycle stage of the product through the above steps, the multi-dimensional feature matrix and the current life cycle stage of the product can be input into a multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current life cycle stage.
[0057] In one embodiment of the present invention, the multi-task learning model may include a shared layer, a task-specific layer, and an output layer. Furthermore, in one embodiment of the present invention, the method for inputting a multi-dimensional feature matrix and the current lifecycle stage of a product into the multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current lifecycle stage may include the following steps: Step 1051: The shared layer uses a Transformer encoder to extract deep correlation features from the multi-dimensional feature matrix; Step 1052: The task-specific layer processes the deep correlation features and the current lifecycle stage of the product to obtain the product preference score and the type and parameters of the operation strategy. Step 1053: The output layer normalizes the product preference scores and operational strategy types and parameters to generate personalized product recommendation schemes and operational strategies adapted to the current lifecycle stage.
[0058] In one embodiment of the present invention, a Transformer encoder is deployed inside the shared layer, which uses a self-attention mechanism to perform deep encoding on the input multi-dimensional feature matrix, calculates the association weights among the product, user and scenario, extracts deep association features across tasks and outputs them to the task-specific layer.
[0059] Furthermore, in one embodiment of the present invention, the aforementioned task-specific layer may include a recommendation task sub-network and an operation strategy task sub-network. In one embodiment of the present invention, the method by which the aforementioned task-specific layer processes deep correlation features and the current lifecycle stage of the product to obtain product preference scores and operation strategy types and parameters may include: inputting deep correlation features into a fully connected neural network in the recommendation task sub-network to obtain a product preference score; and inputting the product preference score and the current lifecycle stage of the product into a second CNN network in the operation strategy task sub-network to obtain an operation strategy type and parameters adapted to the current lifecycle stage of the product.
[0060] In one embodiment of the present invention, the recommendation task subnetwork can receive the deep correlation features output by the shared layer, perform nonlinear mapping through a fully connected neural network, calculate the user's preference score for candidate products, and select the top N products according to the score from high to low, generating a recommended product list containing product ID, product name, and recommendation reasons combined with the life cycle stage description. Simultaneously, based on the scenario time sequence characteristics, the recommendation format is adapted to online product links or offline store locations and placement areas, and the recommendation frequency is dynamically adjusted according to the product's life cycle stage. The operation strategy task subnetwork simultaneously receives the deep correlation features output by the shared layer and the current life cycle stage of the product as input, extracts stage adaptation features through the second CNN network in the operation strategy task subnetwork, and outputs specific operation strategy types and parameters. Specifically, when identified as the introductory stage, the output strategy parameters are to increase promotional investment and set new user discounts; when identified as the growth stage, the output strategy parameters are to optimize recommendation weights and launch bundled packages; when identified as the maturity stage, the output strategy parameters are to optimize pricing and expand sales channels; and when identified as the decline stage, the output strategy parameters are to reduce prices for clearance and decrease recommendation frequency.
[0061] In one embodiment of the present invention, the output layer normalizes the product preference scores and operational strategy types and parameters to generate personalized product recommendation schemes and operational strategies adapted to the current life cycle stage.
[0062] This invention discloses an AI-based recommendation method based on the product lifecycle. It acquires product data through multiple data collection channels, standardizes this data to obtain a standardized multi-source fusion data set, which includes product end-to-end data, user demand data, and scenario-based time-series data. Product lifecycle features are extracted from the end-to-end data, and a fusion neural network model is used to perform spatial and temporal correlation analysis on these features to obtain a product lifecycle feature vector. This vector is then input into a lifecycle stage identification model to determine the current lifecycle stage of the product. User demand features and scenario-based time-series features are extracted from the user demand data and the scenario-based time-series data. The product lifecycle feature vector, user demand features, and scenario-based time-series features are then dimensionally aligned and fused using tensors to obtain a multi-dimensional feature matrix. Finally, the multi-dimensional feature matrix and the current lifecycle stage of the product are input into a multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current lifecycle stage. This invention accurately identifies product lifecycle stages by collecting product data and uses a multi-task learning model to collaboratively generate personalized product recommendation schemes and operational strategies for the current lifecycle stage. This solves the problems of single data dimensions and lagging stage determination in traditional methods, thereby improving the accuracy of product recommendations.
[0063] To achieve the above embodiments, such as Figure 2 As shown in the diagram, this embodiment also provides a structural schematic of an AI recommendation device based on the product lifecycle, which includes: The data acquisition module 201 is used to acquire commodity data through multiple data acquisition channels, and to perform standardized processing on the commodity data to obtain a standardized multi-source fusion data set, which includes commodity full-link data, user demand data and scenario time series data. The feature extraction module 202 is used to extract features from the full-chain data of the product to obtain product lifecycle features, and to use a fusion neural network model to perform spatial and temporal correlation analysis on the product lifecycle features to obtain product lifecycle feature vectors. The first data processing module 203 is used to input the product life cycle feature vector into the life cycle stage identification model to obtain the current life cycle stage of the product. The second data processing module 204 is used to extract features from user demand data and scenario time series data to obtain user demand features and scenario time series features, and to perform dimension alignment and tensor fusion processing on product life cycle feature vector, user demand features and scenario time series features to obtain a multi-dimensional feature matrix. The generation module 205 is used to input the multi-dimensional feature matrix and the current life cycle stage of the product into the multi-task learning model to generate personalized product recommendation schemes and operational strategies adapted to the current life cycle stage.
[0064] In one embodiment of the present invention, the data acquisition module 201 described above is specifically used for: Fill in the missing values in the product collection data to obtain the filled data; Perform anomaly detection and removal on the filled data to obtain the data after removing outliers; The data after removing outliers is subjected to unitization processing to obtain normalized data. Sensitive information in the normalized data is encrypted to obtain desensitized data. Data belonging to the same dimension or the same product entity in the anonymized data are integrated to obtain a standardized multi-source fusion data set.
[0065] In one embodiment of the present invention, the above-mentioned fused neural network model includes a first CNN network, an LSTM network, and an attention mechanism; the above-mentioned feature extraction module 202 is specifically used for: Product lifecycle features are obtained by extracting features from the entire product chain data. These features include transaction features, inventory features, operational features, and environmental features. The transaction features, inventory features, operational features, and environmental features are aligned and concatenated according to their dimensions to obtain an initial feature vector; The initial feature vector is input into the first CNN network to extract the spatial correlation between features, thus obtaining the first feature vector; The first feature vector is input into the LSTM network to extract the temporal correlation of features, thus obtaining the second feature vector; By assigning weights to the first and second feature vectors using an attention mechanism, the product lifecycle feature vector is obtained.
[0066] In one embodiment of the present invention, the second data processing module 204 is specifically used for: Extract user demand keywords from user demand features using TF-IDF algorithm or collaborative filtering algorithm, and transform user demand keywords into user demand feature vectors; The temporal trend features in the temporal features of the scene are extracted by time series analysis algorithm, and the temporal trend features are transformed into a scene temporal feature vector; The Min-Max normalization method is used to transform the product lifecycle feature vector, user demand feature vector, and scenario time series feature vector into the zero-to-one interval and align the sample number with the time dimension to obtain the dimension-aligned product lifecycle feature vector, user demand feature vector, and scenario time series feature vector. By employing tensor fusion, the dimension-aligned product lifecycle feature vector, user demand feature vector, and scenario temporal feature vector are fused to construct a multi-dimensional feature matrix.
[0067] In one embodiment of the present invention, the multi-task learning model includes a shared layer, a task-specific layer, and an output layer; the generation module 205 is specifically used for: The shared layer uses a Transformer encoder to extract deep correlation features from the multi-dimensional feature matrix; The task-specific layer processes deep correlation features and the current lifecycle stage of the product to obtain product preference scores and operational strategy types and parameters. The output layer normalizes product preference scores and operational strategy types and parameters to generate personalized product recommendation schemes and operational strategies adapted to the current lifecycle stage.
[0068] In one embodiment of the present invention, the task-specific layer includes a recommendation task sub-network and an operation strategy task sub-network; the generation module 205 is further configured to: The deep correlation features are input into the fully connected neural network in the recommendation task subnetwork to obtain the product preference score; By inputting the product preference score and the current lifecycle stage of the product into the second CNN network in the operation strategy task subnetwork, the operation strategy type and parameters adapted to the current lifecycle stage of the product are obtained.
[0069] According to an embodiment of the present invention, an AI recommendation device based on the product lifecycle acquires product data through multiple data collection channels, standardizes the product data to obtain a standardized multi-source fusion data set, wherein the multi-source fusion data set includes product end-to-end data, user demand data, and scenario time-series data; extracts features from the product end-to-end data to obtain product lifecycle features, and uses a fusion neural network model to perform spatial and temporal correlation analysis on the product lifecycle features to obtain a product lifecycle feature vector; inputs the product lifecycle feature vector into a lifecycle stage identification model to obtain the current lifecycle stage of the product; extracts features from the user demand data and scenario time-series data to obtain user demand features and scenario time-series features, and performs dimensional alignment and tensor fusion processing on the product lifecycle feature vector, user demand features, and scenario time-series features to obtain a multi-dimensional feature matrix; inputs the multi-dimensional feature matrix and the current lifecycle stage of the product into a multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current lifecycle stage. This invention accurately identifies product lifecycle stages by collecting product data and uses a multi-task learning model to collaboratively generate personalized product recommendation schemes and operational strategies for the current lifecycle stage. This solves the problems of single data dimensions and lagging stage determination in traditional methods, thereby improving the accuracy of product recommendations.
[0070] In the description of this specification, the 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 present 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. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0071] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. An AI-powered recommendation method based on the product lifecycle, characterized in that, include: Product data is collected through multiple data collection channels, and the product data is standardized to obtain a standardized multi-source fusion data set, which includes product end-to-end data, user demand data, and scenario time-series data. Product lifecycle features are obtained by extracting features from the entire product lifecycle data, and spatial and temporal correlation analysis is performed on the product lifecycle features using a fusion neural network model to obtain a product lifecycle feature vector. The product lifecycle feature vector is input into the lifecycle stage identification model to obtain the current lifecycle stage of the product; Feature extraction is performed on the user demand data and the scenario time series data to obtain user demand features and scenario time series features. The product life cycle feature vector, the user demand features and the scenario time series features are then subjected to dimension alignment and tensor fusion processing to obtain a multi-dimensional feature matrix. The multi-dimensional feature matrix and the current lifecycle stage of the product are input into a multi-task learning model to generate personalized product recommendation schemes and operational strategies adapted to the current lifecycle stage.
2. The method according to claim 1, characterized in that, The standardization process of the collected commodity data to obtain a standardized multi-source fusion data set includes: The missing values in the product collection data are filled in to obtain the filled data; The filled data is subjected to anomaly detection and removal to obtain data after removing outliers; The data after removing outliers is subjected to unitization processing to obtain normalized data; The sensitive information in the normalized data is encrypted to obtain the desensitized data; The data belonging to the same dimension or the same product entity in the de-identified data are integrated to obtain a standardized multi-source fusion data set.
3. The method according to claim 1, characterized in that, The fusion neural network model includes a first CNN network, an LSTM network, and an attention mechanism; the product lifecycle features are obtained by extracting features from the entire product lifecycle data, and the fusion neural network model is used to perform spatial and temporal correlation analysis on the product lifecycle features to obtain a product lifecycle feature vector, including: Product lifecycle features are obtained by extracting features from the entire product chain data, wherein the product lifecycle features include transaction features, inventory features, operational features and environmental features; The transaction features, inventory features, operational features, and environmental features are dimensionally aligned and concatenated to obtain an initial feature vector; The initial feature vector is input into the first CNN network to extract the spatial correlation between features, thus obtaining the first feature vector; The first feature vector is input into the LSTM network to extract the temporal correlation of features, thus obtaining the second feature vector; The first feature vector and the second feature vector are weighted by the attention mechanism to obtain the product lifecycle feature vector.
4. The method according to claim 1, characterized in that, The process of dimensional alignment and tensor fusion of the product lifecycle feature vector, the user demand features, and the scenario temporal features yields a multi-dimensional feature matrix, including: The user demand keywords are extracted from the user demand features using the TF-IDF algorithm or collaborative filtering algorithm, and the user demand keywords are transformed into user demand feature vectors. The temporal trend features in the temporal features of the scene are extracted by time series analysis algorithm, and the temporal trend features are transformed into a scene temporal feature vector; The Min-Max normalization method is used to transform the product lifecycle feature vector, the user demand feature vector, and the scenario time series feature vector into a zero-to-one interval and align the sample number with the time dimension to obtain the dimension-aligned product lifecycle feature vector, user demand feature vector, and scenario time series feature vector. The product lifecycle feature vector, user demand feature vector, and scenario time-series feature vector, which are aligned according to the dimensions, are fused to construct a multi-dimensional feature matrix.
5. The method according to claim 1, characterized in that, The multi-task learning model includes a shared layer, a task-specific layer, and an output layer; the step of inputting the multi-dimensional feature matrix and the current lifecycle stage of the product into the multi-task learning model to generate a personalized product recommendation scheme and an operational strategy adapted to the current lifecycle stage includes: The shared layer uses a Transformer encoder to extract deep correlation features from the multi-dimensional feature matrix; The task-specific layer processes the deep correlation features and the current lifecycle stage of the product to obtain product preference scores and operational strategy types and parameters; The output layer normalizes the product preference scores and the operational strategy types and parameters to generate personalized product recommendation schemes and operational strategies adapted to the current lifecycle stage.
6. The method according to claim 5, characterized in that, The task-specific layer includes a recommendation task sub-network and an operational strategy task sub-network; the task-specific layer processes the deep correlation features and the current lifecycle stage of the product to obtain product preference scores and operational strategy types and parameters, including: The deep association features are input into the fully connected neural network in the recommendation task subnetwork to obtain the product preference score; The product preference score and the current lifecycle stage of the product are input into the second CNN network in the operation strategy task sub-network to obtain the operation strategy type and parameters that are adapted to the current lifecycle stage of the product.
7. An AI recommendation device based on the product lifecycle, characterized in that, include: The data acquisition module is used to acquire product data through multiple data acquisition channels, and to perform standardized processing on the product data to obtain a standardized multi-source fusion data set, wherein the multi-source fusion data set includes product full-link data, user demand data and scenario time-series data; The feature extraction module is used to extract features from the full-chain data of the product to obtain product lifecycle features, and to use a fusion neural network model to perform spatial and temporal correlation analysis on the product lifecycle features to obtain a product lifecycle feature vector. The first data processing module is used to input the product lifecycle feature vector into the lifecycle stage identification model to obtain the current lifecycle stage of the product. The second data processing module is used to extract features from the user demand data and the scenario time series data to obtain user demand features and scenario time series features, and to perform dimension alignment and tensor fusion processing on the product life cycle feature vector, the user demand features and the scenario time series features to obtain a multi-dimensional feature matrix. The generation module is used to input the multi-dimensional feature matrix and the current life cycle stage of the product into the multi-task learning model to generate personalized product recommendation schemes and operational strategies adapted to the current life cycle stage.
8. The apparatus according to claim 7, characterized in that, The data acquisition module is specifically used for: The missing values in the product collection data are filled in to obtain the filled data; The filled data is subjected to anomaly detection and removal to obtain data after removing outliers; The data after removing outliers is subjected to unitization processing to obtain normalized data; The sensitive information in the normalized data is encrypted to obtain the desensitized data; The data belonging to the same dimension or the same product entity in the de-identified data are integrated to obtain a standardized multi-source fusion data set.
9. An electronic device, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6.
10. A computer storage medium, wherein, The computer storage medium stores computer-executable instructions; when executed by a processor, the computer-executable instructions can implement the method as described in any one of claims 1-6.