An intelligent matching short video goods recommendation method and management system

Abstract

The application discloses a kind of short video goods recommendation method and management system of intelligent matching, it is related to short video goods recommendation technical field. Including information collection: obtain user information, user information includes: browse information, like unit, past purchase commodity information and user basic information.The application can obtain recommendation information set by the information extraction method, derivative method and past purchase commodity information set, and can obtain relevant goods video set by recommendation information set and goods video library, and can judge what type of video user likes by the setting judging method according to the latest real-time browse information of user, and can obtain goods video of user favorite type by cooperating relevant goods video set according to user favorite type, and by recommending goods video of user favorite type for user, it can improve user experience while improving the transaction efficiency of commodity, so that user and merchant benefit.

Description

Technical Field

[0001] This invention relates to the field of short video e-commerce recommendation technology, specifically to an intelligent matching short video e-commerce recommendation method and management system. Background Technology

[0002] Short video e-commerce refers to a form of product display and sales through short video platforms. In short video e-commerce, hosts or internet celebrities demonstrate the usage effects, features, and advantages of products by posting short videos or live streams, guiding users to make purchases. Short video e-commerce has become a new promotional method in the e-commerce industry, favored by more and more merchants and consumers. Through short video e-commerce, merchants can quickly expand product exposure and increase sales, while users can also learn about product information and gain shopping inspiration through videos, improving the convenience and enjoyment of shopping.

[0003] Intelligent matching recommendations can present users with products that better meet their needs, increasing their willingness to buy and conversion rates, which is beneficial to improving sales and merchants' profitability. Intelligent matching recommendations can reduce the time and energy users spend browsing various product information, thus improving their shopping efficiency.

[0004] Currently, short video e-commerce typically involves recommending live-streaming e-commerce to users. Short video e-commerce also includes video e-commerce, which can be further divided into various types, such as live-streaming segment e-commerce and product review e-commerce. Because live-streaming e-commerce spans a long period and often includes unnecessary interspersed processes, for users who don't have time or enjoy watching live streams, it's necessary to cater to their preferences by recommending e-commerce videos that they like. This improves the user experience, thereby increasing the effectiveness of e-commerce and benefiting both users and merchants. This invention is proposed based on this principle. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent matching method and management system for recommending products through short videos, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent matching short video e-commerce recommendations, comprising:

[0007] Information collection: Obtain user information, including browsing information, likes, past purchase information, and basic user information. Browsing information includes browsing units and browsing completion rate. Obtain the product information library and product information library.

[0008] Demand Acquisition: Based on browsing information, multi-level user preference types and reference browsing units are obtained through judgment methods. Based on like units and reference browsing units, potential user needs are obtained through information extraction methods. The parts of potential user needs that are the same as those of previously purchased products are removed to obtain demand information. Based on previously purchased product information, derivative product information is obtained through derivation methods. The demand information and derivative product information are integrated to obtain a recommendation information set.

[0009] Video Acquisition: By traversing the product-selling video library through the recommendation information set, relevant product-selling video sets are obtained. The product-selling videos that are consistent with the user's preference type in the relevant product-selling video sets are extracted to obtain the multi-level recommendation set.

[0010] Auxiliary information generation: Auxiliary information is obtained based on the recommendation information set through information generation methods. The auxiliary information is then embedded into the corresponding product promotion videos in the multi-level recommendation set to obtain the complete recommendation set.

[0011] Video recommendation: The video recommendation frequency is preset. Based on the user's basic information, the recommendation frequency is combined with the complete recommendation set to recommend videos for product promotion. The behavior information of users watching videos for product promotion is obtained, and the video recommendation frequency is optimized based on the behavior information through adjustment methods.

[0012] The judgment method includes: setting a preset time limit, classifying browsing information based on the time limit to obtain past browsing information and real-time browsing information, real-time browsing information including real-time browsing units and their real-time browsing completion rate, setting a preset threshold for browsing completion rate, removing real-time browsing units with a real-time browsing completion rate less than the threshold browsing completion rate to obtain reference browsing units, classifying several reference browsing units based on their types to obtain several control sets, obtaining user preference types based on the types of reference browsing units in the control sets, and classifying the control sets using a classification method to obtain multi-level user preference types.

[0013] Furthermore, the classification method includes: presetting a hierarchical numerical range, comparing the number of reference browsing units in several control sets with the hierarchical numerical range to obtain correlation, obtaining multi-level user preference types based on the correlation and the type of reference browsing units in the control sets, wherein the correlation feedback information is that when the number of reference browsing units in the control set is greater than the hierarchical numerical range, the type of the control set is a first-level user preference type; when the correlation feedback information is that the number of reference browsing units in the control set is within the hierarchical numerical range, the type of the control set is a second-level user preference type; when the correlation feedback information is that the number of reference browsing units in the control set is less than the hierarchical numerical range, the type of the control set is a third-level user preference type; and integrating the first-level user preference type, the second-level user preference type, and the third-level user preference type in sequence to obtain a multi-level user preference type.

[0014] Furthermore, the information extraction method includes: constructing a user profile based on likes and reference browsing units; obtaining text and image information from likes and reference browsing units to obtain a dataset; extracting product information from the dataset, where product information represents potential user needs; constructing subsequent user profiles based on subsequent likes and reference browsing units; and processing subsequent user profiles using a comparison method to obtain potential user needs.

[0015] Furthermore, the comparison method includes: merging user profiles to obtain a user profile set, merging product information corresponding to the user profiles to obtain a product information set, establishing a correlation between the user profile set and the product information set, setting a threshold overlap, determining the overlap between future user profiles and the user profile set, comparing the threshold overlap with the overlap to obtain correlation, when the correlation feedback information is that there is correlation, then extracting product information from the product information set based on the correlation to obtain potential user needs, when the correlation feedback information is that there is no correlation, then obtaining text information and image information from future likes and future reference browsing units to obtain a future dataset, extracting future product information from the future dataset to obtain potential user needs, and integrating future user profiles and future product information into the user profile set and the product information set respectively.

[0016] Furthermore, the derivation method includes: extracting product features from the product information database to obtain a feature database; splitting previously purchased product information to obtain several products; extracting product features to obtain feature points; combining feature points to obtain feature combinations; integrating feature combinations and feature points to obtain a feature set; matching features associated with the feature set in the feature database to obtain several target features; performing data cleaning on the several target features to obtain a target feature set; and extracting product information from the product information database based on the target feature set to obtain derived product information.

[0017] Furthermore, the information generation method includes: splitting the recommendation information set to obtain product information, retrieving product reviews online to obtain product review information, filtering product review information to obtain positive and negative reviews, and integrating positive and negative reviews to obtain auxiliary information.

[0018] Furthermore, the adjustment method includes: determining whether a user likes the recommended product-selling video based on the user's behavioral information, including browsing completion rate and liking behavior; when the browsing completion rate exceeds a threshold browsing completion rate and there is a liking behavior, the video recommendation frequency is increased; when the browsing completion rate does not exceed the threshold browsing completion rate and there is no liking behavior, the video recommendation frequency is decreased; when the behavioral information is otherwise, the video recommendation frequency is not changed.

[0019] A smart matching short video e-commerce recommendation management system uses the aforementioned smart matching short video e-commerce recommendation method.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] This intelligent matching short video e-commerce recommendation and management system, through set information extraction and derivation methods and past purchase information, can obtain a set of recommendation information. By combining the recommendation information set with an e-commerce video library, relevant e-commerce video sets can be obtained. Through set judgment methods, it can determine what type of video a user prefers based on the user's latest real-time browsing information. Based on the user's preferred type, and by combining relevant e-commerce video sets, e-commerce videos of the user's preferred type can be obtained. By recommending e-commerce videos of the user's preferred type, the system can improve the user experience and increase the efficiency of product transactions, benefiting both users and merchants.

[0022] Meanwhile, in the information extraction method, user profiles are constructed using the user's "like" unit and reference browsing unit. Product information is obtained based on the "like" unit and reference browsing unit, and the user profiles are recorded to obtain a user profile set and a product information set. The user profile set and product information set are continuously enriched during the information extraction process. This allows for the subsequent acquisition of potential user needs by simply comparing the user profile constructed based on the "like" unit and reference browsing unit with the user profile set, and then obtaining the corresponding potential user needs from the product information set through correlation. This improves work efficiency and reduces workload. Through the setting of derivative methods, product information related to the user's past purchase information can be obtained by analyzing the user's past purchase information, which helps to enrich the information in the recommendation information set, making it easier to meet user needs and improve the user experience.

[0023] Meanwhile, by setting the information generation method, auxiliary information for product information, namely positive and negative reviews, can be obtained based on the user's recommendation information set. By allowing users to browse the auxiliary information, users can gain a better understanding of the product, which helps improve the user experience. By setting the adjustment method, the frequency of video recommendations can be adjusted based on user feedback behavior information, increasing the time users spend on the platform and their stickiness, which helps improve the platform's user activity and retention rate, and avoids user churn due to excessive product recommendation videos. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process structure of the upper part of the present invention;

[0025] Figure 2 This is a schematic diagram of the overall process structure of the lower half of the present invention;

[0026] Figure 3This is a schematic diagram of the structure of the derivative method of the present invention;

[0027] Figure 4 This is a schematic diagram of the classification method of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Through intelligent matching recommendations, users can receive more personalized shopping recommendations on short video platforms, increasing the time users spend on the platform and their stickiness, thereby improving user activity and retention rates. Intelligent matching recommendations for short video e-commerce can enhance user experience, increase sales conversion rates, save users time and energy, and improve platform stickiness, which is beneficial to both merchants and users.

[0030] like Figures 1-4 As shown, the present invention provides a technical solution: a method and management system for intelligent matching short video e-commerce recommendations, the method comprising:

[0031] Information collection: Obtain user information, including browsing information, likes, past purchase information, and basic user information. Browsing information includes browsing units and browsing completion rate. Obtain the product information library and product information library.

[0032] Demand Acquisition: Based on browsing information, multi-level user preference types and reference browsing units are obtained through judgment methods. Based on like units and reference browsing units, potential user needs are obtained through information extraction methods. The parts of potential user needs that are the same as those of previously purchased products are removed to obtain demand information. Based on previously purchased product information, derivative product information is obtained through derivation methods. The demand information and derivative product information are integrated to obtain a recommendation information set.

[0033] Video Acquisition: By traversing the product-selling video library through the recommendation information set, relevant product-selling video sets are obtained. The product-selling videos that are consistent with the user's preference type in the relevant product-selling video sets are extracted to obtain the multi-level recommendation set.

[0034] Auxiliary information generation: Auxiliary information is obtained based on the recommendation information set through information generation methods. The auxiliary information is then embedded into the corresponding product promotion videos in the multi-level recommendation set to obtain the complete recommendation set.

[0035] Video recommendation: The video recommendation frequency is preset. Based on the user's basic information, the recommendation frequency is combined with the complete recommendation set to recommend videos for product promotion. The behavior information of users watching videos for product promotion is obtained, and the video recommendation frequency is optimized based on the behavior information through adjustment methods.

[0036] It is important to note that during the information collection phase, user information is acquired through the platform. User privacy must be protected during this process to prevent information leakage. The product and e-commerce video library and product information library are also acquired through the platform. During the demand acquisition phase, multi-level user preference types are obtained by combining browsing information with judgment methods. Video information related to "likes" is also included in the browsing information. By removing content identical to previously purchased products from users' potential needs, the appearance of e-commerce videos for previously purchased products can be avoided, improving the accuracy of product recommendation videos. Through the setting of derivative methods, more related product information can be obtained based on previously purchased product information to generate derivative product information, increasing the richness of the recommendation information set. During the video acquisition phase, the information in the recommendation information set is all... The process involves traversing the product information and video library to search for videos that match the product information. All videos are then aggregated to form a relevant product video set. This traversal can be accomplished using a trained model. This model takes product information as input and outputs product videos. To improve the accuracy of the retrieved videos, the model's output can be manually reviewed. The process of extracting videos from the relevant product video set that match the user's preferred categories to create a multi-level recommendation set can also be done using a combination of a trained model and manual review. User preferred categories are video types, such as educational, comedic, and emotional videos. The video types in the relevant product video set can be obtained from the platform. The multi-level recommendation system sorts the product-selling videos in a manner consistent with the multi-level user preference types. During the auxiliary information generation stage, auxiliary information is generated through information generation methods to help users understand detailed product information and improve the user experience. During video recommendation, product-selling videos are recommended to users based on their basic information and recommendation frequency. When a product-selling video appears, auxiliary information corresponding to the products in the video is output to the user in the form of documents, images, etc., after the video finishes playing, facilitating further understanding of the products in the video. The method provided in this application is embedded in short video e-commerce platforms. The specific short video e-commerce platform is selected based on actual usage needs. This application obtains user browsing information, likes, and past purchase information. When obtaining product information and basic user information, the method is based on the short video e-commerce platform used for integration. The platform's database identifies users based on their registered mobile phone number, WeChat ID, QQ number, and other basic information, and retrieves their browsing information, likes, past purchases, and other basic user information. Regarding the protection of user privacy, the method provided in this application is embedded within the short video e-commerce platform. The implementation involves data transmission within the platform. Since the platform's data transmission operations are inherently encrypted, user privacy is protected during both data acquisition and transmission. Alternatively, custom encryption methods can be used to further protect user privacy.Common encryption methods include symmetric and asymmetric encryption algorithms. These methods generate keys to encrypt and decrypt transmitted data, protecting user privacy. The information extraction method specifically utilizes existing visual recognition technology to identify text and image information within "like" and "reference" units to obtain a dataset. Product information is then extracted from this dataset, and potential user needs are determined based on this product information. The specific process of visual recognition technology involves: collecting relevant data via web scraping (using web crawling techniques) or manual methods, gathering complete information including images and text, manually annotating the collected images to identify "like" units, references, and related text areas. Annotation tools (such as LabelImg and VGG Image) can also be used. Annotators are used to create labeled datasets. Image sizes are adjusted to match the input requirements of the recognition model, such as the commonly used 224x224 pixels. Image denoising, enhancement, and rotation correction are performed to ensure optimal image quality. Text regions in the images are segmented to extract the text to be recognized. The text content is standardized, including removing special characters and converting to lowercase. Pre-trained convolutional neural network (CNN) models, such as ResNet, VGG, and MobileNet, are used to extract image features. A suitable OCR tool (such as Tesseract or EasyOCR) is selected to recognize the text information in the image. The processed image is input into the OCR tool to recognize the text information and output the recognition results. The recognized text and image data are cleaned to remove duplicate, erroneous, or useless data to obtain the dataset. After obtaining the dataset, the extracted text and image information can be verified through manual review or by using a validation dataset to ensure the accuracy of the data. Product information is extracted from the dataset. The process involves obtaining a comparison set by acquiring the names and keywords of products sold on short video platforms. Combining this comparison set with existing data, the similarities between the two sets are used to obtain product information. Past purchase information is then removed from this product information to obtain demand information, which consists of several product names or keywords. This demand information is further derived using derivation methods and combined with the original product information to obtain a recommendation information set. Auxiliary information involves splitting the recommendation information set to obtain several product listings, retrieving positive and negative reviews from the platform, and integrating these to obtain auxiliary information. The complete recommendation set consists of the short videos to be recommended and the reviews of the products featured in those videos. Multi-level user preference types are determined by classifying user-preferred video types based on likes and viewing time. Recommendation frequency and time limits are set based on actual usage. The default recommendation frequency is 3 short videos per hour, and the time limit is a specific time point, such as six months prior to the current time.Specialized models refer to deep learning models, such as convolutional neural networks. The correlation is obtained by comparing the number of reference browsing units and the hierarchical numerical ranges in several control sets, i.e., the quantitative relationship.

[0037] The judgment method includes: setting a preset time limit, classifying browsing information based on the time limit to obtain past browsing information and real-time browsing information, real-time browsing information including real-time browsing units and their real-time browsing completion rate, setting a preset threshold for browsing completion rate, removing real-time browsing units with a real-time browsing completion rate less than the threshold browsing completion rate to obtain reference browsing units, classifying several reference browsing units based on their types to obtain several control sets, obtaining user preference types based on the types of reference browsing units in the control sets, and classifying the control sets using a classification method to obtain multi-level user preference types.

[0038] It's important to note that user preferences can change over time. The time limit should be determined based on the actual situation, ideally as close to real-time as possible while ensuring sufficient browsing data samples. The threshold browsing completion rate represents the user's completion rate for a given browsing unit (video). Under normal circumstances, the threshold browsing completion rate can be set to 80%. The type of the reference browsing unit can be obtained through the video tags on the platform. By setting the judgment method, it determines the user's preferred video type based on the latest real-time browsing information, avoiding information gaps and preventing the recommendation of videos based on outdated preferences, thus improving the user experience. Furthermore, by categorizing user preferences, it can prioritize recommending videos with the highest user preferences for product promotion, improving sales efficiency and enhancing the user experience.

[0039] like Figure 4 As shown, the classification method includes: pre-setting a hierarchical numerical range, comparing the number of reference browsing units in several control sets with the hierarchical numerical range to obtain the correlation, and obtaining multi-level user preference types based on the correlation and the type of reference browsing units in the control sets. The feedback information of the correlation is that when the number of reference browsing units in the control set is greater than the hierarchical numerical range, the type of the control set is a first-level user preference type; when the number of reference browsing units in the control set is within the hierarchical numerical range, the type of the control set is a second-level user preference type; when the number of reference browsing units in the control set is less than the hierarchical numerical range, the type of the control set is a third-level user preference type; and the first-level user preference type, second-level user preference type, and third-level user preference type are integrated in sequence to obtain a multi-level user preference type.

[0040] It's important to note that the tiered numerical range is determined based on the actual situation. The tiered numerical range represents the range of reference browsing units within the comparison set. In multi-level user preference types, the preference types are ordered as Level 1, Level 2, and Level 3. For example, if the tiered numerical range is 4-6, then if the number of reference browsing units in the comparison set exceeds 6, the comparison set type is Level 1; if the number of reference browsing units is 4, 5, or 6, the comparison set type is Level 2; and if the number of reference browsing units is less than 4, the comparison set type is Level 3. This tiered processing of user preference types facilitates subsequent video recommendations, allowing for sequential recommendations of promotional videos and improving the user experience.

[0041] like Figure 1 As shown, the information extraction method includes: constructing user profiles based on likes and reference browsing units; obtaining text and image information from likes and reference browsing units to obtain a dataset; extracting product information from the dataset, which represents potential user needs; constructing subsequent user profiles based on subsequent likes and reference browsing units; and processing these subsequent user profiles using a comparison method to obtain potential user needs. The comparison method includes: merging user profiles to obtain a user profile set; merging product information corresponding to the user profiles to obtain a product information set; and establishing a relationship between the user profile set and the product information set. A preset threshold overlap is set to determine the overlap between the future user profile and the user profile set. The overlap between the threshold overlap and the overlap is compared to obtain the correlation. When the correlation feedback is that there is a correlation, the product information in the corresponding product information set is extracted to obtain the user's potential needs based on the correlation. When the correlation feedback is that there is no correlation, the text information and image information in the future like unit and the future reference browsing unit are obtained to obtain the future dataset. The future product information in the future dataset is extracted to obtain the user's potential needs, and the future user profile and future product information are integrated into the corresponding user profile set and product information set.

[0042] It's important to note that textual and image information within "like" units and reference browsing units can be obtained using existing information retrieval and visual image technologies. Product information can be extracted from the dataset by training a specialized extraction model. Building user profiles serves as a reference. Subsequent "like" units and reference browsing units refer to those processed later. The threshold overlap is determined based on the specific circumstances. For example, in the first instance of building user profiles and obtaining product information, the first user profile is a user profile set, and the first set of product information is a product information set. In the second instance of building user profiles, the second user profile is compared with the user profiles in the user profile set. The overlap is judged against a threshold overlap to determine the correlation. When the overlap exceeds the threshold overlap, the feedback information indicating correlation is relevant. The second user profile is constructed using product information from the first user profile. If the overlap threshold is not reached, the feedback information indicates no relevance. Then, text and image information from the "like" and "reference browsing" units are obtained to create a dataset. Product information is extracted from this dataset, and the second user profile is integrated into the user profile set, while the product information from the second user profile is integrated into the product information set. This process is repeated for the third user profile construction. Through the information extraction method, the user profile set and product information set are continuously enriched during the process. This allows for the subsequent acquisition of potential user needs by simply comparing the user profile constructed from the "like" and "reference browsing" units with the user profile set, and then using the correlation to obtain the corresponding potential user needs from the product information set. This improves work efficiency and reduces workload.

[0043] like Figure 2 and Figure 3 As shown, the derivative method includes: extracting product features from the product information database to obtain a feature database; splitting previously purchased product information to obtain several products; extracting product features to obtain feature points; combining feature points to obtain feature combinations; integrating feature combinations and feature points to obtain a feature set; matching features associated with the feature set in the feature database to obtain several target features; processing the several target features using data processing methods to obtain a target feature set; and extracting product information from the product information database based on the target feature set to obtain derived product information.

[0044] It's important to note that extracting product features can be done by training a specialized model. The product's features are its name, and the process involves combining feature points to obtain feature combinations. For example, with three feature points 1, 2, and 3, the combinations are 1 and 2, 2 and 3, 1 and 3, and 1, 2, and 3. Matching the feature set to features in the feature library involves sequentially selecting feature combinations or feature points from the feature set and traversing the feature library to obtain associated target features. This matching process can be performed by training a specialized model. For example, if the feature combination is "dishwasher" and "tableware," the target features obtained after traversing the feature library would be detergent, chopsticks, etc., associated with dishwashers and tableware. When a feature combination cannot be matched in the feature library, the next feature combination or feature point is matched in turn. Data processing methods are used to clean the target features to ensure their quality. By setting derivative methods, product information related to previously purchased products can be obtained based on user's past purchase information, enriching the recommendation information set and better meeting user needs and improving the user experience.

[0045] like Figure 2 As shown, the information generation method includes: splitting the recommendation information set to obtain product information, retrieving product reviews online to obtain product review information, filtering product review information to obtain positive and negative reviews, and integrating positive and negative reviews to obtain auxiliary information.

[0046] It's important to note that the process of obtaining product reviews through online retrieval can be achieved manually by searching for product information on shopping platforms or search engines. Filtering positive and negative reviews is a current technology. High-quality positive and negative reviews can be selected based on word count criteria. This filtering process can be conducted using trained professional models or manually by staff. The review process can employ sampling to improve efficiency, or it can utilize trained deep learning models to review the output results. By setting the information generation method, positive and negative reviews can be obtained, making it easier for users to browse and improving the user experience.

[0047] like Figure 2 As shown, the adjustment method includes: judging whether users like the recommended product-selling videos based on user behavior information, including browsing completion rate and liking behavior. When the browsing completion rate exceeds the threshold and there is a liking behavior, the video recommendation frequency is increased. When the browsing completion rate does not exceed the threshold and there is no liking behavior, the video recommendation frequency is decreased. When the behavior information is otherwise, the video recommendation frequency is not changed.

[0048] It is important to note that by adjusting the settings, the frequency of video recommendations can be adjusted based on user feedback, increasing the time users spend on the platform and their engagement. This helps improve user activity and retention, and prevents users from leaving the platform due to an excessive number of product-related video recommendations.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. An intelligent matching short video goods recommendation method, characterized in that: The method comprises: information collection: obtaining user information, including browsing information, like units, past purchase commodity information and user basic information, browsing information including browsing units and their browsing completion, obtaining goods video library and commodity information library; demand acquisition: based on browsing information, the multi-level user preference type and the reference browsing unit are obtained by a judgment method, based on the like unit and the reference browsing unit, the user's potential demand is obtained by an information extraction method, the same part of the user's potential demand as the past purchase commodity information is removed to obtain the demand information, the derivative commodity information is obtained based on the past purchase commodity information by a derivation method, and the recommendation information set is obtained by integrating the demand information and the derivative commodity information; video acquisition: traversing the goods video library through the recommendation information set to obtain the related goods video set, and extracting the goods video in the related goods video set consistent with the multi-level user preference type to obtain the multi-level recommendation set; assistant information generation: based on the recommendation information set, the auxiliary information is obtained by an information generation method, and the auxiliary information is embedded in the corresponding goods video in the multi-level recommendation set to obtain the complete recommendation set; video recommendation: presetting the video recommendation frequency, based on the user basic information, the video recommendation frequency is adjusted based on the behavior information to optimize the video recommendation frequency; The judgment method comprises: presetting a time limit, classifying the browsing information based on the time limit to obtain past browsing information and real-time browsing information, the real-time browsing information including real-time browsing units and their real-time browsing completion, a threshold browsing completion, removing the real-time browsing units with real-time browsing completion less than the threshold browsing completion to obtain the reference browsing units, classifying a plurality of reference browsing units based on the type of the reference browsing units to obtain a plurality of comparison sets, obtaining the user preference type based on the type of the reference browsing units in the comparison set, and classifying the comparison set by a classification method to obtain the multi-level user preference type.

2. The intelligent matching short video goods recommendation method according to claim 1, characterized in that: The classification method comprises: presetting a grading numerical interval, comparing the number of reference browsing units in the comparison set with the grading numerical interval to obtain the relevance, obtaining the multi-level user preference type based on the relevance and the type of the reference browsing units in the comparison set, the feedback information of the relevance is that the number of reference browsing units in the comparison set is greater than the grading numerical interval, the type of the comparison set is the first-level user preference type, the feedback information of the relevance is that the number of reference browsing units in the comparison set is located in the grading numerical interval, the type of the comparison set is the second-level user preference type, and the feedback information of the relevance is that the number of reference browsing units in the comparison set is less than the grading numerical interval, the type of the comparison set is the third-level user preference type.

3. The intelligent matching short video goods recommendation method according to claim 1, characterized in that: The information extraction method comprises the following steps: constructing a user portrait based on a like unit and a reference browsing unit, obtaining a data set by obtaining script information and image information in the like unit and the reference browsing unit, extracting commodity information in the data set, the commodity information being a potential demand of the user, constructing a subsequent user portrait based on a subsequent like unit and a subsequent reference browsing unit, and obtaining the potential demand of the user by processing the subsequent user portrait through a comparison method.

4. The intelligent matching short video goods recommendation method according to claim 3, characterized in that: The comparison method comprises the following steps: merging the user portraits to obtain a user portrait set, merging commodity information corresponding to the user portraits to obtain a commodity information set, establishing an association relationship between the user portrait set and the commodity information set, presetting a threshold coincidence degree, judging a coincidence degree in the user portrait set and the subsequent user portrait, comparing the threshold coincidence degree and the coincidence degree to obtain relevance, when feedback information of the relevance is relevant, extracting commodity information in the commodity information set based on the association relationship to obtain the potential demand of the user, when the feedback information of the relevance is irrelevant, obtaining subsequent data set by obtaining script information and image information in the subsequent like unit and the subsequent reference browsing unit, extracting subsequent commodity information in the subsequent data set to obtain the potential demand of the user, and integrating the subsequent user portrait and the subsequent commodity information into the user portrait set and the commodity information set.

5. The intelligent matching short video goods recommendation method according to claim 1, characterized in that: The derivation method comprises the following steps: extracting a commodity feature in a commodity information library to obtain a feature library, splitting previous purchase commodity information to obtain a plurality of commodities, extracting a feature of the commodity to obtain a feature point, combining the feature points to obtain a feature combination, integrating the feature combination and the feature point to obtain a feature set, matching a feature associated with the feature set in the feature library to obtain a plurality of target features, obtaining a target feature set after data cleaning of the plurality of target features, and extracting commodity information in the commodity information library based on the target feature set to obtain derived commodity information.

6. The intelligent matching short video goods recommendation method according to claim 1, characterized in that: The information generation method comprises the following steps: splitting a recommendation information set to obtain commodity information, online searching evaluation of the commodity information to obtain commodity evaluation information, screening the commodity evaluation information to obtain high-quality positive evaluation information and high-quality negative evaluation information, and integrating the high-quality positive evaluation information and the high-quality negative evaluation information to obtain auxiliary information.

7. The intelligent matching short video goods recommendation method according to claim 1, characterized in that: The adjustment method comprises the following steps: judging whether the user likes a recommended live streaming video according to behavior information of the user, the behavior information comprising a browsing completion degree and a like behavior, when the browsing completion degree in the behavior information exceeds a threshold browsing completion degree and the like behavior exists, increasing a video recommendation frequency, when the browsing completion degree in the behavior information does not exceed the threshold browsing completion degree and the like behavior does not exist, decreasing the video recommendation frequency, and when the behavior information is in other conditions, not changing the video recommendation frequency.

8. An intelligent matching short video goods recommendation management system, characterized in that: A short video live streaming recommendation method with intelligent matching is used.

Images