An AI-based commodity recommendation method and system

Abstract

The application discloses an AI-based commodity recommendation method and system, and relates to the technical field of artificial intelligence e-commerce recommendation. The method collects multi-dimensional behavior data and attribute data of users and completes preprocessing, relies on a multi-fusion AI model that fuses a Transformer network, collaborative filtering and content feature matching, mines long-term and short-term implicit interests of users and explicit features of commodities, and generates personalized interest labels; high-matching-degree commodities are screened through multi-dimensional weighted scoring to form an initial candidate pool, and the candidate pool is optimized in combination with time, region, festival, activity and other scene information to generate a final recommended list that is scenarized and diversified; meanwhile, model parameters and recommendation strategies are iteratively optimized based on real-time feedback of users to realize closed-loop dynamic optimization. The application solves the defects of traditional recommendation techniques, such as inaccurate interest mining, poor scene adaptability, homogeneous recommendation and inability to dynamically iterate, effectively improves commodity recommendation accuracy and conversion rate, and is suitable for personalized recommendation requirements of various e-commerce platforms, and has strong technical universality and landing property.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, big data processing, and e-commerce recommendation technology, specifically to an AI-based product recommendation method and system. Background Technology

[0002] With the rapid development of e-commerce and online retail, the volume of goods on platforms is growing exponentially, and user consumption demands are becoming increasingly personalized and diversified. Accurate, efficient, and dynamic product recommendation technology has become a core key for e-commerce platforms to improve user experience, increase product conversion rates, and boost platform revenue. Currently, the mainstream product recommendation technologies in the industry are mainly divided into three categories: collaborative filtering-based recommendation algorithms, content feature matching-based recommendation algorithms, and traditional machine learning-based recommendation schemes.

[0003] Traditional collaborative filtering algorithms primarily rely on historical user interactions to uncover user-product associations, without depending on product attribute features. While possessing some group matching capabilities, these algorithms suffer from a severe cold-start problem, failing to provide effective recommendations for new users and products. They also exhibit drawbacks such as data sparsity, highly homogenized recommendation results, and an inability to capture dynamic user interests. Content feature-based matching algorithms, relying on inherent product attributes for matching and recommendation, can address the cold-start problem for new products. However, they can only uncover explicit user preferences, failing to capture implicit or latent consumer needs, resulting in low personalization.

[0004] Existing AI recommendation solutions based on traditional machine learning often employ single neural network models for feature extraction. These models have weak feature fusion capabilities, failing to effectively distinguish between users' long-term stable interests and short-term transient interests. Furthermore, they generally ignore the influence of contextual factors on users' purchasing decisions, resulting in static and fixed recommendation results that cannot adapt to dynamic changes in time, holidays, regions, platform activities, etc. In addition, existing technologies lack a robust user feedback iteration mechanism, failing to dynamically optimize recommendation strategies based on real-time user interactions. This leads to low recommendation accuracy and conversion rates, making it difficult to meet the current demands for refined and intelligent e-commerce recommendations.

[0005] In summary, existing product recommendation technologies suffer from several technical shortcomings, including incomplete interest mining, poor scenario adaptability, severe homogenization, poor cold start performance, and inability to dynamically iterate and optimize. There is an urgent need for a high-precision, highly adaptable, and dynamically iterative AI-powered intelligent product recommendation solution to address these issues. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing an AI-based product recommendation method and system. This system solves the problems of inaccurate user interest mining, poor scenario adaptability, homogenized recommendations, inability to dynamically iterate, and poor cold start performance in existing recommendation technologies. It enables personalized, scenario-based, and dynamic intelligent product recommendations, significantly improving recommendation accuracy and product conversion rates. Technical solution

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an AI-based product recommendation method, comprising the following steps: S1. Real-time collection of multi-dimensional behavioral data and basic attribute data of target users. The multi-dimensional behavioral data includes, but is not limited to, user browsing behavior, click behavior, favorite behavior, add-to-cart behavior, order payment behavior, search keywords, page dwell time and interaction frequency. The basic attribute data includes user age, gender, region, consumption level and historical consumption preferences. The collected raw data is cleaned, deduplicated, missing value imputation and standardized preprocessing are performed to remove abnormal interference data and obtain a standardized user behavior dataset. S2. Input the standardized user behavior dataset into the pre-trained multi-fusion AI recommendation model. The multi-fusion AI recommendation model integrates Transformer feature extraction network, collaborative filtering algorithm and content feature matching algorithm. The Transformer feature extraction network mines the user's long-term and short-term implicit interest features. The collaborative filtering algorithm matches similar user groups and similar product association features. The content feature matching algorithm extracts explicit features such as product category, price, style and attributes. The three types of features are integrated to generate a personalized interest tag system for users. S3. Based on the user's personalized interest tag system, perform feature matching and weighted scoring on the products in the full product library, and select a preset number of highly matched products according to the scoring results to generate an initial product recommendation candidate pool. S4. Obtain current scenario information, including time scenario, geographical scenario, holiday hotspots, platform activities and real-time user demand scenario. Based on the scenario information, perform weight fine-tuning, deduplication filtering and supplementation of unpopular high-quality products in the initial product recommendation candidate pool to generate a final product recommendation list adapted to the current scenario. S5. Push the final product recommendation list to the user terminal for display, and at the same time collect user feedback data on the recommended products in real time. Based on the user feedback data, iteratively optimize the model parameters of the multi-fusion AI recommendation model, and dynamically update user interest tags and recommendation strategies.

[0008] Furthermore, the pre-training process of the multi-fusion AI recommendation model in step S2 includes: S201. Construct a training dataset, which includes historical user behavior data, full product feature data, scenario data, and corresponding user click conversion tags. S202. Build a basic model framework, with the Transformer encoder as the core feature extraction module, and build user interest feature branches, product feature branches, and scene feature branches respectively. S203. A weighted attention mechanism is introduced to assign weights to the features of each branch, thereby strengthening the influence of highly correlated features and weakening the interference of invalid features. S204. The gradient descent algorithm is used to iteratively train the model, with recommendation accuracy, click-through rate and conversion rate as the core evaluation indicators. The model parameters are optimized until the model converges, and a pre-trained multi-fusion AI recommendation model is obtained.

[0009] Furthermore, the specific method for feature matching and weighted scoring of the goods described in step S3 is as follows: S301. Match user interest tags with the attribute tags of each product one by one, and calculate the feature matching score. S302. Generate a product quality score by combining the product's historical sales volume, positive review rate, listing duration, and merchant reputation. S303. Generate a user suitability score based on the user's historical consumption price range and consumption frequency; S304. The feature matching score, product quality score, and user suitability score are weighted and summed according to the preset weight ratio to obtain the comprehensive recommendation score for each product.

[0010] Furthermore, the optimization of the initial product recommendation candidate pool based on scenario information in step S4 includes: Identify the current time, holiday, and platform marketing campaign scenarios to increase the recommendation weight of essential products in the corresponding scenarios; match the consumption habits and best-selling products in the user's region to replace products with low regional compatibility; while retaining popular products, implant niche and high-quality new products in a preset proportion to avoid the problem of homogenized recommended content.

[0011] Furthermore, the model iterative optimization process described in step S5 includes: The system provides real-time statistics on clicks, add-to-cart counts, sales, and skip rates for recommended products, distinguishing between positive and negative feedback data. Based on positive feedback data, it strengthens the matching weight of corresponding interest tags and product features, while based on negative feedback data, it reduces the weight of corresponding features or removes related product types. The system also updates model parameters in batches on a regular basis to achieve dynamic iterative optimization of recommendation performance.

[0012] Secondly, the present invention provides an AI-based product recommendation system for executing the above-mentioned AI-based product recommendation method. The system includes a data acquisition and preprocessing module, an AI intelligent modeling module, a product matching and scoring module, a scenario-optimized recommendation module, and a recommendation push and iterative optimization module. The data acquisition and preprocessing module is used to collect multidimensional user behavior data and basic attribute data in real time, complete data cleaning and standardization preprocessing, and output a standardized user behavior dataset. The AI ​​intelligent modeling module has a built-in pre-trained multi-fusion AI recommendation model, which is used to receive standardized user behavior datasets, mine users' long-term and short-term interest features, and generate a personalized user interest tag system. The product matching and scoring module is used to perform feature matching and comprehensive scoring on all products based on users' personalized interest tags, and to filter and generate an initial product recommendation candidate pool. The scenario optimization and recommendation module is used to obtain real-time scenario information, optimize and adjust the initial product recommendation candidate pool, and generate the final product recommendation list. The recommendation push and iterative optimization module is used to push the recommendation list to the terminal, collect user feedback data and drive the AI ​​model to iteratively optimize and dynamically update the recommendation strategy.

[0013] Furthermore, the AI ​​intelligent modeling module includes a feature extraction unit, a fusion matching unit, and a label generation unit; The feature extraction unit adopts a Transformer network structure to extract users' long-term stable interest features and short-term real-time behavioral interest features. The fusion matching unit is used to fuse collaborative filtering algorithms and content feature matching algorithms to complete the association matching analysis between users and between users and products. The tag generation unit is used to integrate various features to generate a structured and quantifiable personalized interest tag system for users.

[0014] Furthermore, the scene optimization recommendation module includes a scene recognition unit, a weight adjustment unit, and a content optimization unit; The scene recognition unit is used to identify time, region, festival, activity, and real-time user demand scene information in real time. The weight adjustment unit is used to dynamically adjust the product rating weight and interest tag weight according to different scenario types; The content optimization unit is used to perform deduplication of recommended products, removal of inferior products, addition of new products, and homogenization optimization.

[0015] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the steps of the above-mentioned AI-based product recommendation method.

[0016] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the above-described AI-based product recommendation method. Beneficial effects

[0017] Compared with the prior art, the present invention has the following advantages: 1. This invention adopts an AI model that integrates multiple algorithms such as Transformer network, collaborative filtering, and content feature matching. It can simultaneously mine users' long-term and short-term implicit interests and explicit product features, solving the problems of incomplete interest mining and weak feature extraction in traditional single algorithms. It significantly improves the accuracy of user preference recognition and effectively improves the cold start effect for new users and new products.

[0018] 2. This invention introduces a multi-dimensional scene recognition mechanism, which can adjust the recommendation weight and content in real time according to dynamic scenes such as time, region, festival, and platform activities. This breaks the defects of the static and fixed traditional recommendation scheme, realizes the scene-based adaptation of recommendation results, and fits the real-time consumption needs of users.

[0019] 3. This invention comprehensively considers user preferences, product quality, and user suitability through a multi-dimensional weighted scoring mechanism, while also adding new products and niche high-quality products as supplementary strategies. This effectively solves the problems of severe homogenization in traditional recommendations and the monopoly of recommendation positions by best-selling products, thus enriching the diversity of recommendation content.

[0020] 4. This invention sets up a closed-loop user feedback iteration mechanism, which can dynamically optimize model parameters and recommendation strategies in real time based on user interaction feedback, so as to achieve continuous iteration and upgrading of recommendation effect, adapt to changes in user interests and market consumption trends, and the long-term recommendation accuracy, product click-through rate and conversion rate are significantly better than traditional solutions.

[0021] 5. The overall technical solution of this invention has a clear architecture and a high degree of modularity. It can be adapted to various e-commerce, retail, and content-driven sales platforms for recommendation scenarios. It has strong versatility and scalability, and has extremely high practical application value and commercialization value. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the overall process of the AI-based product recommendation method in this embodiment of the invention.

[0024] Figure 2 This is a structural block diagram of an AI-based product recommendation system in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention. Example 1: Example of an AI-based product recommendation method

[0026] like Figure 1 As shown in the figure, this embodiment discloses an AI-based product recommendation method, and the specific execution steps are as follows: The first step is data collection and preprocessing. The system captures all interaction data of target users on the e-commerce platform in real time, including historical behavior data such as browsing, favorites, and orders over the past 30 days, as well as short-term real-time behavior data such as searches, instantaneous clicks, and page dwell time on the current day. It also collects basic attribute data such as user gender, age group, residential region, and historical average order value. Standardized preprocessing operations are performed on the collected raw data to remove interfering data such as bot clicks, invalid refreshes, and abnormal redirects, fill in missing user behavior data, standardize data format and units, and generate a standardized, structured user behavior dataset, providing accurate data support for subsequent feature extraction.

[0027] The second step involves feature extraction and tag generation from a multi-fusion AI model. The pre-processed dataset is input into a pre-trained multi-fusion AI recommendation model. The model uses the multi-head attention mechanism of the Transformer encoder to accurately capture users' long-term stable consumption preferences and short-term sudden changes in interest. Simultaneously, it combines a collaborative filtering algorithm to match similar user groups within the platform with the same demographics and consumption preferences, extracting common consumption characteristics of these groups. A content feature matching algorithm is used to analyze the explicit attributes of all products on the platform, such as category, price range, style, material, and applicable scenarios. The model performs weighted fusion of these three types of features, removes redundant features, and generates a multi-dimensional structured interest tag system that includes user-preferred categories, price sensitivity, style preferences, and scenario requirements.

[0028] The third step is weighted product scoring and candidate pool generation. For the platform's full product database, each product attribute tag is precisely matched with user interest tags to calculate a feature matching score of 0-10. Combining the product's sales volume over the past 7 days, user positive review rate, merchant qualification level, and new product weight, a product quality score of 0-10 is generated. Combining the user's historical purchase price range and purchase frequency, a user consumption suitability score is calculated. The three scores are weighted and summed according to a 5:3:2 weighting ratio to obtain a comprehensive product recommendation score. The top 50 products with the highest scores are then selected to generate the initial recommendation candidate pool.

[0029] The fourth step involves scenario-based optimization to generate the final recommendation list. The system identifies current scenario information in real time, such as holiday scenarios, seasonal changes, platform promotional activities, and user location scenarios. For example, in the high-temperature summer scenario, the weight of sun protection, cooling, and heat-relieving products is increased; in the Double Eleven promotion scenario, products with promotional activities and large discounts are prioritized; for users in southern regions, products suitable for the southern climate and local customs are prioritized. At the same time, the candidate pool of products is deduplicated, and low-quality, low-rated products are removed. Popular products and high-quality new products are retained in a 9:1 ratio to avoid homogenization of recommendations, ultimately generating a final recommendation list containing 20 products.

[0030] Step 5: Push Display and Model Iteration Optimization. The final recommendation list, sorted by rating, is pushed to user homepages, "You May Also Like," pop-up recommendations, and other terminal locations. Real-time data is collected on user feedback regarding recommended products, including clicks, adding to cart, placing orders, skipping, and blocking. Clicks, adding to cart, and placing orders are marked as positive feedback, while quickly skipping or multiple instances of not clicking are marked as negative feedback. Based on the feedback data, the model feature weights are adjusted in real-time, strengthening the weight of categories of high user interest and reducing the weight of ineffective product features. Model parameters are iterated and updated daily in batches to achieve dynamic optimization of the recommendation strategy. Example 2: Example of an AI-based product recommendation system

[0031] like Figure 2 As shown, this embodiment is used to execute the recommended method described in Embodiment 1. The system as a whole adopts a modular architecture, and its specific components and functions are as follows: 1. Data Acquisition and Preprocessing Module: Equipped with a real-time data capture interface and data cleaning algorithm, it collects multi-dimensional user behavior and attribute data around the clock, completes deduplication, noise reduction, and standardization processing, and outputs structured datasets to ensure the accuracy and validity of input data.

[0032] 2. AI Intelligent Modeling Module: This module includes a feature extraction unit, a fusion matching unit, and a tag generation unit. The feature extraction unit uses the Transformer network to mine users' long-term and short-term interest features; the fusion matching unit combines two algorithms to complete the matching of users and products; and the tag generation unit integrates multi-dimensional features and outputs standardized user interest tags to provide a basis for product matching.

[0033] 3. Product Matching and Scoring Module: Built-in multi-dimensional weighted scoring algorithm to batch complete the matching and scoring of all products with user interests, quickly filter high-match products, build an initial recommendation candidate pool, and ensure the basic adaptability of recommended products.

[0034] 4. Scene-Optimized Recommendation Module: The scene recognition unit perceives scene information such as time, region, and activity in real time, the weight adjustment unit dynamically optimizes the product rating weight, and the content optimization unit completes deduplication, elimination of inferior products, and supplementation of new products, so as to realize the scene-based and diversified optimization of recommendation results.

[0035] 5. Recommendation Push and Iterative Optimization Module: Responsible for the terminal distribution and display of the recommendation list, synchronously collecting user feedback data, establishing a feedback database, driving the iterative update of model parameters, and forming a closed-loop operation mechanism of "data collection - intelligent recommendation - feedback optimization".

Claims

1. An AI-based product recommendation method, characterized in that, Includes the following steps: S1. Real-time collection of multi-dimensional behavioral data and basic attribute data of target users. The multi-dimensional behavioral data includes user browsing behavior, click behavior, collection behavior, add-to-cart behavior, order payment behavior, search keywords, page dwell time and interaction frequency. The basic attribute data includes user age, gender, region, consumption level and historical consumption preferences. The collected raw data is cleaned, deduplicated, missing value imputation and standardized preprocessing are performed to remove abnormal interference data and obtain a standardized user behavior dataset. S2. Input the standardized user behavior dataset into the pre-trained multi-fusion AI recommendation model. The multi-fusion AI recommendation model integrates Transformer feature extraction network, collaborative filtering algorithm and content feature matching algorithm. The Transformer feature extraction network mines the user's long-term and short-term implicit interest features. The collaborative filtering algorithm matches similar user groups and similar product association features. The content feature matching algorithm extracts explicit features such as product category, price, style and attributes. The three types of features are integrated to generate a personalized interest tag system for users. S3. Based on the user's personalized interest tag system, perform feature matching and weighted scoring on the products in the full product library, and select a preset number of highly matched products according to the scoring results to generate an initial product recommendation candidate pool. S4. Obtain current scenario information, including time scenario, geographical scenario, holiday hotspots, platform activities and real-time user demand scenario. Based on the scenario information, perform weight fine-tuning, deduplication filtering and supplementation of unpopular high-quality products in the initial product recommendation candidate pool to generate a final product recommendation list adapted to the current scenario. S5. Push the final product recommendation list to the user terminal for display, and at the same time collect user feedback data on the recommended products in real time. Based on the user feedback data, iteratively optimize the model parameters of the multi-fusion AI recommendation model, and dynamically update user interest tags and recommendation strategies.

2. The AI-based product recommendation method according to claim 1, characterized in that, The pre-training process of the multi-fusion AI recommendation model in step S2 includes: S201. Construct a training dataset, which includes historical user behavior data, full product feature data, scenario data, and corresponding user click conversion tags. S202. Build a basic model framework, with the Transformer encoder as the core feature extraction module, and build user interest feature branches, product feature branches, and scene feature branches respectively. S203. A weighted attention mechanism is introduced to assign weights to the features of each branch, thereby strengthening the influence of highly correlated features and weakening the interference of invalid features. S204. The gradient descent algorithm is used to iteratively train the model, with recommendation accuracy, click-through rate and conversion rate as the core evaluation indicators. The model parameters are optimized until the model converges, and a pre-trained multi-fusion AI recommendation model is obtained.

3. The AI-based product recommendation method according to claim 1, characterized in that, The specific method for feature matching and weighted scoring of the goods described in step S3 is as follows: S301. Match user interest tags with the attribute tags of each product one by one, and calculate the feature matching score. S302. Generate a product quality score by combining the product's historical sales volume, positive review rate, listing duration, and merchant reputation. S303. Generate a user suitability score based on the user's historical consumption price range and consumption frequency; S304. The feature matching score, product quality score, and user suitability score are weighted and summed according to the preset weight ratio to obtain the comprehensive recommendation score for each product.

4. The AI-based product recommendation method according to claim 1, characterized in that, Step S4, which involves optimizing the initial product recommendation candidate pool based on scenario information, includes: Identify the current time, holiday, and platform marketing campaign scenarios to increase the recommendation weight of essential products in the corresponding scenarios; match the consumption habits and best-selling products in the user's region to replace products with low regional compatibility; while retaining popular products, implant niche and high-quality new products in a preset proportion to avoid the problem of homogenized recommended content.

5. The AI-based product recommendation method according to claim 1, characterized in that, The model iterative optimization process described in step S5 includes: The system provides real-time statistics on clicks, add-to-cart counts, sales, and skip rates for recommended products, distinguishing between positive and negative feedback data. Based on positive feedback data, it strengthens the matching weight of corresponding interest tags and product features, while based on negative feedback data, it reduces the weight of corresponding features or removes related product types. The system also updates model parameters in batches on a regular basis to achieve dynamic iterative optimization of recommendation performance.

6. An AI-based product recommendation system, characterized in that, The system is used to execute the AI-based product recommendation method according to any one of claims 1-5, and the system includes a data acquisition and preprocessing module, an AI intelligent modeling module, a product matching and scoring module, a scenario-optimized recommendation module, and a recommendation push and iterative optimization module; The data acquisition and preprocessing module is used to collect multidimensional user behavior data and basic attribute data in real time, complete data cleaning and standardization preprocessing, and output a standardized user behavior dataset. The AI ​​intelligent modeling module has a built-in pre-trained multi-fusion AI recommendation model, which is used to receive standardized user behavior datasets, mine users' long-term and short-term interest features, and generate a personalized user interest tag system. The product matching and scoring module is used to perform feature matching and comprehensive scoring on all products based on users' personalized interest tags, and to filter and generate an initial product recommendation candidate pool. The scenario optimization and recommendation module is used to obtain real-time scenario information, optimize and adjust the initial product recommendation candidate pool, and generate the final product recommendation list. The recommendation push and iterative optimization module is used to push the recommendation list to the terminal, collect user feedback data and drive the AI ​​model to iteratively optimize and dynamically update the recommendation strategy.

7. The AI-based product recommendation system according to claim 6, characterized in that, The AI ​​intelligent modeling module includes a feature extraction unit, a fusion matching unit, and a label generation unit; The feature extraction unit adopts a Transformer network structure to extract users' long-term stable interest features and short-term real-time behavioral interest features. The fusion matching unit is used to fuse collaborative filtering algorithms and content feature matching algorithms to complete the association matching analysis between users and between users and products. The tag generation unit is used to integrate various features to generate a structured and quantifiable personalized interest tag system for users.

8. The AI-based product recommendation system according to claim 6, characterized in that, The scene optimization recommendation module includes a scene recognition unit, a weight adjustment unit, and a content optimization unit; The scene recognition unit is used to identify time, region, festival, activity, and real-time user demand scene information in real time. The weight adjustment unit is used to dynamically adjust the product rating weight and interest tag weight according to different scenario types; The content optimization unit is used to perform deduplication of recommended products, removal of inferior products, addition of new products, and homogenization optimization.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program stored in the memory to implement the steps of the AI-based product recommendation method according to any one of claims 1-5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the AI-based product recommendation method according to any one of claims 1-5.

Images