An e-commerce intelligent recommendation method and system based on artificial intelligence
By reconstructing user behavior sequences across all dimensions, mapping product lifecycle features, and implementing a reinforcement learning decision-making loop, the system addresses the issues of one-sided user characteristics and unbalanced product structure in existing e-commerce intelligent recommendations, achieving accurate matching of user needs and improving platform operational efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DONGYING DONGWANG INTERNET INFORMATION TECH CO LTD
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-03
AI Technical Summary
Existing intelligent recommendation technologies in e-commerce fail to effectively consider the time series and spatial correlation of user behavior data, and do not take into account the changing patterns of products throughout their entire life cycle. This results in one-sided and inaccurate user feature representation, and recommendation decisions are prone to homogenization and imbalance in product structure.
By employing reinforcement learning to reconstruct user behavior sequences across all dimensions, dynamic feature mapping throughout the product lifecycle, and a reinforcement learning-based end-to-end decision-making loop, and combining the temporal and spatial correlations of user behavior data, dynamic adaptation rules for product and user features are established, and an end-to-end recommendation decision-making loop is constructed to iteratively optimize recommendation results in real time.
It has achieved a comprehensive and accurate improvement in user feature representation, truly reflected user consumption needs, optimized product turnover efficiency, solved the problems of user recommendation fatigue and platform product structure imbalance, and achieved simultaneous improvement in user experience and platform operational efficiency.
Smart Images
Figure CN122335409A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of e-commerce, and in particular to an intelligent recommendation method and system for e-commerce based on artificial intelligence. Background Technology
[0002] With the large-scale development of the e-commerce industry, intelligent recommendation has become a core technology for improving user shopping experience and optimizing platform operation efficiency. However, existing e-commerce intelligent recommendation technology still has many technical shortcomings that are difficult to address simultaneously.
[0003] In existing technologies, most recommendation solutions use collaborative filtering and basic deep learning models to achieve product matching. In the user feature processing stage, only explicit behavioral data such as browsing and ordering are collected, without taking into account the time series and scenario space correlation of behavioral data. This makes it impossible to uncover users' implicit consumption needs. At the same time, noisy data and missing nodes in the behavioral sequence are not effectively corrected, resulting in one-sided and inaccurate user feature representation, which cannot truly reflect users' consumption intentions.
[0004] In the product feature processing stage, existing solutions mostly adopt static product attribute features and fail to establish adaptation rules based on the dynamic changes of the entire product life cycle. This makes it impossible to match the operational needs of different stages of product launch, growth, stabilization, and clearance, resulting in a serious lack of adaptation between product features and user needs.
[0005] In the recommendation decision-making process, existing solutions often focus on click-through rate and short-term conversion rate as the sole optimization goals, without taking into account the multi-objective collaboration of long-term user retention and platform product turnover efficiency. Furthermore, the application of reinforcement learning and enhancement learning is often isolated, failing to form a closed loop of feature processing and decision optimization. This easily leads to problems such as homogeneous recommendations and a severe Matthew effect for best-selling products, ultimately causing user recommendation fatigue and an imbalance in the platform's product structure.
[0006] Therefore, it is necessary to propose an intelligent recommendation method and system for e-commerce based on artificial intelligence to solve the above problems. Summary of the Invention
[0007] The main objective of this invention is to provide an intelligent recommendation method and system for e-commerce based on artificial intelligence, which can effectively solve the problems in the background technology.
[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An AI-based intelligent recommendation method for e-commerce includes the following steps: S1: Reinforcement learning-based feature reconstruction of user full-dimensional spatiotemporal behavior sequence, used to collect user behavior data of the whole life cycle and all scenarios on the e-commerce platform, sort it in spatiotemporal dual dimensions to generate the original behavior sequence, and reconstruct the spatiotemporal related behavior feature set by correcting noise and completing implicit related nodes through reinforcement learning. S2: Dynamic adaptation feature mapping of product lifecycle value, used to collect full-dimensional data of product lifecycle on the platform, divide the product lifecycle into stages, establish adaptation mapping rules between product features and user reconstructed feature sets at each stage, eliminate redundant features, and generate a dimension-aligned dynamic feature set of products. S3: Reinforcement learning-based end-to-end recommendation decision closed-loop generation, used to build a reinforcement learning decision model. It takes user end-to-end conversion, long-term retention and platform product turnover as joint reward functions, inputs user and product feature sets, iterates in real time to output adapted recommendation sets, and feeds back to form a full closed-loop linkage. S4: Contextualized and precise delivery of recommendation results. This is used to adjust the sorting priority of the recommended product set according to the user's current context, adapt to the needs of the context, and deliver the corresponding recommendation slots. At the same time, it records all user behavior data after delivery as input for the next iteration.
[0009] Preferably, S1 specifically includes: S101: Collect all user behavior data throughout the entire lifecycle of the e-commerce platform, including browsing, adding to cart, favorites, placing orders, returns and exchanges, comments, dwell time on product detail pages, page scrolling trajectory, search term modification trajectory, and cross-time period behavior interval data. Sort by timestamp and scene space in two dimensions to generate the original behavior sequence. The scene space includes the homepage, search page, product detail page, live room, and shopping cart page. S102: Using a reinforcement learning framework, the semantic integrity and relevance of the behavior sequence are used as reward objectives to correct missing nodes and noisy behaviors in the original behavior sequence. At the same time, implicit associations between user behaviors are mined, implicit feature nodes in the sequence are completed, and a reconstructed spatiotemporal related behavior feature set is generated.
[0010] Preferably, step S2 specifically includes: S202: Collect full lifecycle data of all products on the platform, including listing duration, inventory changes, price fluctuations, evaluation changes, after-sales rate, competition in the same category, and supply chain fulfillment timeliness data, and divide the lifecycle into stages according to the product's new product launch period, growth period, stable period, and clearance period. S203: For each lifecycle stage, establish adaptation mapping rules between product features and user behavior feature sets after reconstruction, eliminate redundant product features that are not related to the current user behavior, and generate a dynamic product feature set aligned with the user feature dimensions.
[0011] Preferably, step S3 specifically includes: S303: Construct a reinforcement learning decision-making model, using user full-link conversion efficiency, long-term retention rate, and platform product turnover efficiency as joint reward functions, and the aforementioned user spatiotemporal related behavior feature set and product dynamic feature set as model input states; S303: The model updates the input state in real time at each user behavior node, iteratively optimizes the recommendation decision, and outputs a set of recommended products that perfectly matches the current user state and product lifecycle stage, avoiding the problems of homogeneous recommendations and over-recommendation of high-frequency best-selling products; S304: User feedback data on the recommendation results is transmitted back in real time to the reinforcement learning feature reconstruction stage in the first step, completing the closed-loop linkage of feature reconstruction, feature mapping, recommendation decision, and feedback iteration.
[0012] Preferably, step S4 specifically includes: S401: Based on the user's current scene space, the output recommended product set is sorted and prioritized. When the user is on the shopping cart page, products that complement the products in the shopping cart are recommended first. When the user is on the search results page, products that match the user's search intent with the highest lifecycle are recommended first. S402: Deploy the adjusted set of recommended products to the corresponding recommendation slots in the scenario, and record all user behavior data after deployment as input for the next round of feature reconstruction and model iteration.
[0013] An AI-based intelligent recommendation system for e-commerce includes a user spatiotemporal behavior feature enhancement and reconstruction module, a product lifecycle dynamic feature mapping module, a reinforcement learning end-to-end recommendation decision module, and a scenario-based recommendation delivery execution module. The user spatiotemporal behavior feature enhancement and reconstruction module specifically includes: The end-to-end behavior acquisition submodule is used to collect user behavior data across multiple scenarios throughout the entire lifecycle and generate time-series sequences. The reinforcement learning feature correction and completion submodule is used to complete noise correction, implicit association mining and feature completion through reinforcement learning, and outputs a spatiotemporal association behavior feature set.
[0014] Preferably, the product lifecycle dynamic feature mapping module specifically includes: Product Lifecycle Segmentation Submodule: Used to complete the collection of full-lifecycle data and the segmentation of lifecycle stages for products; The feature adaptation mapping submodule is used to establish adaptation mapping rules between product and user features, and outputs a dynamic feature set of products with aligned dimensions.
[0015] Preferably, the reinforcement learning end-to-end recommendation decision module specifically includes: Multi-objective reward decision submodule: It takes the feature set of the product lifecycle dynamic feature mapping module as input, and takes user conversion, retention and product turnover as the joint reward output recommendation set; The feedback loop iteration submodule is used to synchronously complete the closed-loop transmission of feedback data.
[0016] Preferably, the scenario-based recommendation delivery execution module specifically includes: Scene priority adaptation submodule: used to complete the scene-based sorting adjustment and precise delivery of the recommendation set; Delivery data feedback submodule: Used to synchronously collect all delivery data for system iteration.
[0017] Compared with existing technologies, this invention provides an intelligent recommendation method and system for e-commerce based on artificial intelligence, which has the following beneficial effects: This AI-based intelligent recommendation method and system for e-commerce reconstructs the features of users' full-dimensional spatiotemporal behavior sequences through reinforcement learning. It comprehensively collects user behavior data across multiple scenarios throughout the entire lifecycle, taking into account the behavioral correlations in both time and space. Simultaneously, it uses the semantic integrity and relevance of the behavior sequence as the reward objective to complete noisy behavior correction, implicit correlation mining, and implicit feature completion. This significantly improves the comprehensiveness and accuracy of user feature representation, truly restoring users' potential consumption needs and providing a highly reliable input foundation for subsequent end-to-end recommendation decisions.
[0018] This AI-based intelligent recommendation method and system for e-commerce dynamically maps the value of a product throughout its entire lifecycle and establishes feature adaptation rules based on the operational characteristics of different stages of the product's lifecycle. This achieves dimensional alignment and precise adaptation between product features and user behavior features after reconstruction, breaking through the limitations of existing static product features. While accurately matching user needs, it also fully considers the operational needs of the entire product lifecycle, effectively improving the platform's product turnover efficiency and optimizing the platform's overall product structure.
[0019] This AI-based intelligent recommendation method and system for e-commerce constructs a closed-loop recommendation decision-making process across the entire value chain through reinforcement learning. Using user conversion efficiency, long-term retention rate, and platform product turnover efficiency as joint reward functions, it incorporates the user spatiotemporal behavioral feature set and product dynamic feature set output by reinforcement learning as real-time input. This enables real-time iterative optimization of recommendation decisions. Simultaneously, user feedback data is fed back to the feature reconstruction stage in real time, achieving deep linkage between feature processing and decision optimization in reinforcement learning. This synergistic and complementary approach addresses the core pain points of existing technologies, such as single-objective optimization and isolated application, effectively achieving a two-way synchronous improvement in user experience and platform operational efficiency. Attached Figure Description
[0020] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.
[0022] Example 1: like Figure 1 As shown, an AI-based intelligent recommendation method for e-commerce includes the following steps: S1: Reinforcement learning-based feature reconstruction of user full-dimensional spatiotemporal behavior sequences. This is used to collect user behavior data across the entire lifecycle and all scenarios on e-commerce platforms, generate original behavior sequences by sorting them in both spatiotemporal dimensions, and then use reinforcement learning to correct noise and complete implicit related nodes to reconstruct a spatiotemporal related behavior feature set. Specifically, it includes: S101: Collect all user behavior data throughout the entire lifecycle of the e-commerce platform, including browsing, adding to cart, favorites, placing orders, returns and exchanges, comments, dwell time on product detail pages, page scrolling trajectory, search term modification trajectory, and cross-time period behavior interval data. Sort by timestamp and scene space in two dimensions to generate the original behavior sequence. The scene space includes the homepage, search page, product detail page, live room, and shopping cart page. S102: Using a reinforcement learning framework, the semantic integrity and relevance of the behavior sequence are used as reward objectives to correct missing nodes and noisy behaviors in the original behavior sequence. At the same time, implicit associations between user behaviors are mined, implicit feature nodes in the sequence are completed, and a reconstructed spatiotemporal related behavior feature set is generated.
[0023] S2: Dynamic adaptation feature mapping for the entire product lifecycle value. This is used to collect full-dimensional data on the entire product lifecycle on the platform. It divides the product lifecycle into stages, establishes adaptation mapping rules between product features at each stage and user-reconstructed feature sets, eliminates redundant features, and generates a dimensionally aligned dynamic feature set for the product. Specifically, it includes: S202: Collect full lifecycle data of all products on the platform, including listing duration, inventory changes, price fluctuations, evaluation changes, after-sales rate, competition in the same category, and supply chain fulfillment timeliness data, and divide the lifecycle into stages according to the product's new product launch period, growth period, stable period, and clearance period. S203: For each lifecycle stage, establish adaptation mapping rules between product features and user behavior feature sets after reconstruction, eliminate redundant product features that are not related to the current user behavior, and generate a dynamic product feature set aligned with the user feature dimensions.
[0024] S3: A closed-loop generation system for end-to-end recommendation decisions based on reinforcement learning. This system constructs a reinforcement learning decision model, using user end-to-end conversion, long-term retention, and platform product turnover as the joint reward function. It takes user and product feature sets as input, it iteratively outputs an adapted recommendation set in real time, and provides feedback to form a fully closed-loop linkage. Specifically, it includes: S303: Construct a reinforcement learning decision-making model, using user full-link conversion efficiency, long-term retention rate, and platform product turnover efficiency as joint reward functions, and the aforementioned user spatiotemporal related behavior feature set and product dynamic feature set as model input states; S303: The model updates the input state in real time at each user behavior node, iteratively optimizes the recommendation decision, and outputs a set of recommended products that perfectly matches the current user state and product lifecycle stage, avoiding the problems of homogeneous recommendations and over-recommendation of high-frequency best-selling products; S304: User feedback data on the recommendation results is transmitted back in real time to the reinforcement learning feature reconstruction stage in the first step, completing the closed-loop linkage of feature reconstruction, feature mapping, recommendation decision, and feedback iteration.
[0025] S4: Contextualized and precise delivery of recommendation results. This function adjusts the sorting priority of the recommended product set based on the user's current context, adapts to the scenario requirements, and delivers recommendations to the corresponding positions. It also records all user behavior data after delivery as input for the next iteration. Specifically, this includes: S401: Based on the user's current scene space, the output recommended product set is sorted and prioritized. When the user is on the shopping cart page, products that complement the products in the shopping cart are recommended first. When the user is on the search results page, products that match the user's search intent with the highest lifecycle are recommended first. S402: Deploy the adjusted set of recommended products to the corresponding recommendation slots in the scenario, and record all user behavior data after deployment as input for the next round of feature reconstruction and model iteration.
[0026] Example 2: An AI-based intelligent recommendation system for e-commerce includes a user spatiotemporal behavior feature enhancement and reconstruction module, a product lifecycle dynamic feature mapping module, a reinforcement learning end-to-end recommendation decision module, and a scenario-based recommendation delivery execution module. The user spatiotemporal behavior feature enhancement and reconstruction module specifically includes: The end-to-end behavior acquisition submodule is used to collect user behavior data across multiple scenarios throughout the entire lifecycle and generate time-series sequences. The reinforcement learning feature correction and completion submodule is used to complete noise correction, implicit association mining and feature completion through reinforcement learning, and outputs a spatiotemporal association behavior feature set.
[0027] The product lifecycle dynamic feature mapping module specifically includes: Product Lifecycle Segmentation Submodule: Used to complete the collection of full-lifecycle data and the segmentation of lifecycle stages for products; The feature adaptation mapping submodule is used to establish adaptation mapping rules between product and user features, and outputs a dynamic feature set of products with aligned dimensions.
[0028] The reinforcement learning end-to-end recommendation decision module specifically includes: Multi-objective reward decision submodule: It takes the feature set of the product lifecycle dynamic feature mapping module as input, and takes user conversion, retention and product turnover as the joint reward output recommendation set; The feedback loop iteration submodule is used to synchronously complete the closed-loop transmission of feedback data.
[0029] The contextualized recommendation and delivery execution module specifically includes: Scene priority adaptation submodule: used to complete the scene-based sorting adjustment and precise delivery of the recommendation set; Delivery data feedback submodule: Used to synchronously collect all delivery data for system iteration.
[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. An artificial intelligence-based electronic commerce intelligent recommendation method, characterized in that: The following steps are included: S1: Reinforcement learning-based feature reconstruction of user full-dimensional spatiotemporal behavior sequence, used to collect user behavior data of the whole life cycle and all scenarios on the e-commerce platform, sort it in spatiotemporal dual dimensions to generate the original behavior sequence, and reconstruct the spatiotemporal related behavior feature set by correcting noise and completing implicit related nodes through reinforcement learning. S2: Dynamic adaptation feature mapping of product lifecycle value, used to collect full-dimensional data of product lifecycle on the platform, divide the product lifecycle into stages, establish adaptation mapping rules between product features and user reconstructed feature sets at each stage, eliminate redundant features, and generate a dimension-aligned dynamic feature set of products. S3: Reinforcement learning-based end-to-end recommendation decision closed-loop generation, used to build a reinforcement learning decision model. It takes user end-to-end conversion, long-term retention and platform product turnover as joint reward functions, inputs user and product feature sets, iterates in real time to output adapted recommendation sets, and feeds back to form a full closed-loop linkage. S4: Contextualized and precise delivery of recommendation results. This is used to adjust the sorting priority of the recommended product set according to the user's current context, adapt to the needs of the context, and deliver the corresponding recommendation slots. At the same time, it records all user behavior data after delivery as input for the next iteration. 2.The AI-based e-commerce intelligent recommendation method of claim 1, wherein: Specifically, S1 includes: S101: Collect all user behavior data throughout the entire lifecycle of the e-commerce platform, including browsing, adding to cart, favorites, placing orders, returns and exchanges, comments, dwell time on product detail pages, page scrolling trajectory, search term modification trajectory, and cross-time period behavior interval data. Sort by timestamp and scene space in two dimensions to generate the original behavior sequence. The scene space includes the homepage, search page, product detail page, live room, and shopping cart page. S102: Using a reinforcement learning framework, the semantic integrity and relevance of the behavior sequence are used as reward objectives to correct missing nodes and noisy behaviors in the original behavior sequence. At the same time, implicit associations between user behaviors are mined, implicit feature nodes in the sequence are completed, and a reconstructed spatiotemporal related behavior feature set is generated. 3.The AI-based e-commerce intelligent recommendation method of claim 2, wherein: Specifically, S2 includes: S202: Collect full lifecycle data of all products on the platform, including listing duration, inventory changes, price fluctuations, evaluation changes, after-sales rate, competition in the same category, and supply chain fulfillment timeliness data, and divide the lifecycle into stages according to the product's new product launch period, growth period, stable period, and clearance period. S203: For each lifecycle stage, establish adaptation mapping rules between product features and user behavior feature sets after reconstruction, eliminate redundant product features that are not related to the current user behavior, and generate a dynamic product feature set aligned with the user feature dimensions. 4.The AI-based e-commerce intelligent recommendation method of claim 3, wherein: Specifically, S3 includes: S303: Construct a reinforcement learning decision-making model, using user full-link conversion efficiency, long-term retention rate, and platform product turnover efficiency as joint reward functions, and the aforementioned user spatiotemporal related behavior feature set and product dynamic feature set as model input states; S303: The model updates the input state in real time at each user behavior node, iteratively optimizes the recommendation decision, and outputs a set of recommended products that perfectly matches the current user state and product lifecycle stage, avoiding the problems of homogeneous recommendations and over-recommendation of high-frequency best-selling products; S304: User feedback data on the recommendation results is transmitted back in real time to the reinforcement learning feature reconstruction stage in the first step, completing the closed-loop linkage of feature reconstruction, feature mapping, recommendation decision, and feedback iteration. 5.The AI-based e-commerce intelligent recommendation method of claim 4, wherein: Specifically, S4 includes: S401: Based on the user's current scene space, the output set of recommended products is sorted and prioritized. When the user is on the shopping cart page, products that complement the products in the shopping cart are recommended first. When the user is on the search results page, products that match the user's search intent with the highest lifecycle are recommended first. S402: Deploy the adjusted set of recommended products to the corresponding recommendation slots in the scenario, and record all user behavior data after deployment as input for the next round of feature reconstruction and model iteration.
6. An artificial intelligence-based e-commerce intelligent recommendation system employing an artificial intelligence-based e-commerce intelligent recommendation method according to any one of claims 1-5, comprising a user spatio-temporal behavior feature enhancement reconstruction module, a commodity full life cycle dynamic feature mapping module, a reinforcement learning full link recommendation decision module, and a scenario-based recommendation delivery execution module, characterized in that: The user spatiotemporal behavior feature enhancement and reconstruction module specifically includes: The end-to-end behavior acquisition submodule is used to collect user behavior data across multiple scenarios throughout the entire lifecycle and generate time-series sequences. The reinforcement learning feature correction and completion submodule is used to complete noise correction, implicit association mining and feature completion through reinforcement learning, and outputs a spatiotemporal association behavior feature set. 7.The AI-based e-commerce intelligent recommendation system according to claim 6, characterized in that: The product lifecycle dynamic feature mapping module specifically includes: Product Lifecycle Segmentation Submodule: Used to complete the collection of full-lifecycle data and the segmentation of lifecycle stages for products; The feature adaptation mapping submodule is used to establish adaptation mapping rules between product and user features, and outputs a dynamic feature set of products with aligned dimensions. 8.The AI-based e-commerce intelligent recommendation system according to claim 6, characterized in that: The reinforcement learning end-to-end recommendation decision module specifically includes: Multi-objective reward decision submodule: It takes the feature set of the product lifecycle dynamic feature mapping module as input, and takes user conversion, retention and product turnover as the joint reward output recommendation set; The feedback loop iteration submodule is used to synchronously complete the closed-loop transmission of feedback data.
9. The intelligent recommendation system for e-commerce based on artificial intelligence according to claim 6, characterized in that: The scenario-based recommendation and delivery execution module specifically includes: Scene priority adaptation submodule: used to complete the scene-based sorting adjustment and precise delivery of the recommendation set; Delivery data feedback submodule: Used to synchronously collect all delivery data for system iteration.