Real-time product recommendation system and its implementation method

Abstract

This provides a real-time product recommendation system and a method for implementing it. [Solution] During implementation, real-time information from the internet is captured at regular intervals (periodically), analyzed by an artificial intelligence model to generate tags, which are then attached to related products. Simultaneously, the corresponding weight values ​​are assigned or updated to the tags. This provides immediacy when recommending products, automatically sorting products that best match consumers' real-time needs, accurately reflecting real-time information in product recommendations, and achieving the highest level of electronic marketing effectiveness.

Description

Technical Field

[0001] The present invention relates to e - marketing, and more particularly to a real - time product recommendation system that uses artificial intelligence to generate relevant tags from network data regularly (at fixed times) to achieve automatic release of products, and a method for implementing the same.

Background Art

[0002] In the e - commerce environment, product recommendation is an essential function. Merchants upload product data to form multiple tags for convenient classification and management, and also introduce products that meet the needs of consumers. Tags are usually configured manually, so the recommendation method cannot fully consider changes in consumers and new needs. Not only can the content of tags not fully meet the search intentions of each customer, but a large amount of time cost is also spent on the product tagging work.

[0003] In conventional patent documents, for example, Patent Document 1 discloses "Automated Categorization of Products in a Merchant Catalog", which captures the metadata of text in product information, estimates the relationship between each product and each category, and searches for the most relevant recommended products based on search words. In addition, in other conventional patent documents, for example, Patent Document 2 discloses "A Recommendation Method for Personalized Internet Marketing", Patent Document 3 discloses "An E - commerce Assistance System for Collecting Product Information Comprehensively and Conducting Sales Forecasts", Patent Document 4 discloses "A Product Recommendation Method, Device, and Computing Equipment", Patent Document 5 discloses "A Product Advertising Distribution Device, a Product Advertising Distribution Method, and a Product Advertising Distribution Control Program", Patent Document 6 discloses "A Product Recommendation System", and Patent Document 7 discloses "A Product Recommendation Method".

Prior Art Documents

Patent Documents

[0004] [Patent Document 1] U.S. Patent No. 10,528,907 [Patent Document 2] Taiwan Patent No. 611362B [Patent Document 3] Taiwan Patent No. 828426B Publication [Patent Document 4] Chinese Patent No. 111582984A [Patent Document 5] Japanese Patent No. 4859892B2 [Patent Document 6] Japanese Patent No. 7353655B2 [Patent Document 7] Japanese Patent Publication No. 2024081552 [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] As mentioned above, numerous methods already exist for capturing keywords and product recommendations in text information. These methods involve calculating weight values ​​based on the frequency of keyword appearance in the text, constructing tag sets using machine learning, and matching them with recommended products. Additionally, tags are generated by creating models using natural language processing and image recognition. However, using only text information of products as the source for tag generation has limited effectiveness in meeting consumers' actual purchasing intentions and real-time demand. In particular, products related to currently trending topics often had to be manually organized and the product list updated. Furthermore, in knowledge base articles with limited time, recommended products frequently encountered problems such as expired or discontinued items, requiring manual updates to display new related products. Therefore, the challenges awaiting a solution were how to improve the limitations on recommended products in product recommendation systems, how to accurately reflect real-time information in product recommendations, and how to immediately meet consumer demand.

[0006] This invention was made through the inventors' diligent research in view of the above-mentioned problems, and its purpose is to provide a real-time product recommendation system and a method for implementing it, which periodically captures network data, extracts words, generates related tags and their weights using artificial intelligence, and tags corresponding related products. [Means for solving the problem]

[0007] To achieve the above objective, a real-time product recommendation system and its implementation method according to a certain aspect of the present invention include a real-time product recommendation system in which the real-time product recommendation system periodically captures real-time information on the internet within a time interval (e.g., within 6 hours) using a real-time capture module, transmits it to a word extraction module for sorting, and extracts real-time tags. Next, a tag combining module generates combined tags and attaches them to related products. Finally, a recommended product list is constructed that is updated in real time as needed, responding to the needs of different consumers in real time.

[0008] Other features of the present invention will be made clearer by description in this specification and the accompanying drawings. [Brief explanation of the drawing]

[0009] [Figure 1] This is a diagram showing a real-time product recommendation system according to one embodiment of the present invention. [Figure 2] This is a flowchart (1) showing a method for implementing a real-time product recommendation system according to one embodiment of the present invention. [Figure 3] This is a flowchart (2) showing a method for implementing a real-time product recommendation system according to one embodiment of the present invention. [Figure 4] This is a schematic diagram (1) illustrating the implementation of the present invention. [Figure 5] This is a schematic diagram (2) illustrating the implementation of the present invention. [Figure 6] This is a schematic diagram (3) illustrating the implementation of the present invention. [Figure 7] It is a flowchart (III) for explaining the method of the present invention. [Figure 8] It is a schematic diagram (IV) for explaining the implementation of the present invention. [Figure 9] It is a schematic diagram (V) for explaining the implementation of the present invention. [Figure 10] It is a schematic diagram (VI) for explaining the implementation of the present invention. [Figure 11] It is a schematic diagram (VII) for explaining the implementation of the present invention. [Figure 12] It is a schematic diagram (VIII) for explaining the implementation of the present invention. [Figure 13] It is a schematic diagram (IX) for explaining the implementation of the present invention. [Figure 14] It is a schematic diagram (X) for explaining the implementation of the present invention.

Embodiments for Carrying out the Invention

[0010] Hereinafter, embodiments of the present invention will be described in detail. However, the present invention is not limited thereto, and various modifications are possible within the described scope. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention.

[0011] As shown in Figure 1, the real-time product recommendation system 1 according to the present invention is connected (communicatively connected) to the Internet 2 and the user's information device 3, and can be a Web server for conducting organized sales. The so-called commercial transaction site may be one of the following, or a combination thereof, such as a shopping site, funding site, event site, or trading site. The real-time product recommendation system 1 according to the present invention supports functions such as network connection, data storage, database management, and application execution. The server may be independent physical hardware or a virtualized virtual server. The real-time product recommendation system 1 includes a central processing module 11, and a storage module 12, a real-time capture module 13, a word extraction module 14, a tag joining module 15, and a product display module 16 are connected (communicatively connected) to the central processing module 11. In this specification, one site is given as an example, but it may be extended to multiple sites, which will be explained in detail below.

[0012] (1) The central processing module 11 drives each unit to which information is connected (connected in a communicative manner), operates each module, and operates the server. The central processing module 11 has at least functions such as logical operations, temporary storage of operation results, and storage of execution instruction positions, and may be one of the following or a combination thereof: "Central Processing Unit, CPU", "Graphics Processing Unit, GPU", "Micro Processing Unit, MPU", or "Micro Control Unit, MCU", and the present invention is not limited to these. (2) The storage module 12 is used to store various information and data, and may be, for example, one of "Solid State Disk or Solid State Drive, SSD", (Hard Disk Drive, HDD), "Static Random Access Memory, SRAM", "Dynamic Random Access Memory, DRAM", or a combination thereof, or "Cloud Drive". The storage module 12 includes a learning database 121, a tag database 122, and a product database 123. The learning database 121 mainly stores existing information, existing related information, and keyword models, and the keyword models are configured by performing machine learning using existing information and existing related information. The tag database 122 mainly stores a plurality of existing tags and real-time tags generated by the tag combination module 15. The product database 123 stores a plurality of products (product information), and each product information is set to be related to one or more existing tags and existing weights corresponding to the existing tags, respectively. (3) The real-time capture module 13 acquires at least one piece of real-time information and its corresponding related information via the Internet 2. The real-time capture module 13 acquires real-time information and related information mainly using methods such as browser search, data capture, "Web Crawler," or a combination thereof, but the present invention is not limited to these. The real-time information refers to textual information such as online news, forum articles, SNS articles, product data, and publicly available documents, or a combination thereof, but the present invention is not limited to this. For example, it may be in media format such as audio, images, or videos. Furthermore, the real-time information may include not only one natural language or one natural language system, but also multiple natural languages ​​or pidgin languages. The related information may include text that links various numerical values ​​such as the number of browses, the number of link clicks, the number of quotes, the number of product purchases, the frequency of word occurrences, the frequency of reach, the frequency of clicks, the frequency of common word occurrences, user behavior data, relationships with external sites, expert knowledge conversion, probability space, information entropy (Shannon Entropy), spatial distribution, etc., and their conversions, with weights associated with other important words, or one type or combination of related data of words within text, and the present invention is not limited to these. The user behavior data is, for example, an action event performed by the user within the site using the information device 3, and may include browse records, click records, stay records, shopping cart records, purchase records, etc., and the above user behavior data is weighted based on its relationship. For example, the weight of a purchase record may be set to 5, a record of adding items to a shopping cart to 3, a click record to 2, and a browse record to 1. The relevance to external sites is determined by analyzing and calculating data crawled from external sources, determining the related words for each, and assigning a score to the crawled data text.The aforementioned conversion of expert knowledge refers to converting current in-house expert rules and existing knowledge into quantified numerical values ​​such as explicit knowledge, but the present invention is not limited to this. (4) The word extraction module 14 performs natural language processing (NLP) on real-time information and extracts at least one real-time word. The real-time word is a hot word such as one with a high search frequency or topical relevance, or a characteristic word such as one with a high relevance to a product or one with functionality, and generates at least one real-time tag using a keyword model. For training and learning this keyword model, "Machine Learning" which allows for self-training and learning such as "Unsupervised learning," "Self-Supervised Learning," and "Heuristic Algorithms" is mainly used, but the present invention is not limited to these. Existing information that is autonomously collected by the real-time capture module 13 or stored in the learning database 121, along with existing related information, is used as training data. The real-time tags generated by the word extraction module 14 are further subjected to vector operations to generate tag feature vectors, and then vector operations are performed on user behavior data to generate user feature vectors. Finally, the similarity between the tag feature vectors and user feature vectors is calculated to generate tag weights, which are used in subsequent related calculations. Furthermore, real-time tags are not limited to a single language or language system; they may be generated using the same or similar words from different languages ​​or language systems. (5) The tag combining module 15 receives the real-time tags generated by the word extraction module 14 and generates at least one combined tag. The combined tag includes a catch, or the catch is combined with at least one sub-catch, and the real-time tag is used as the catch. The combined tag model performs machine learning, deep learning, and artificial intelligence-related technologies such as "Large Language Model, LLM" to list multiple combined tags that have different sub-catches. The catch and sub-catch are further subjected to vector operations to form a combined tag vector matrix consisting of a catch feature vector and a sub-catch feature vector, respectively, and the relationship and relevance of the catch and sub-catch in different combined tags are calculated. Subsequently, similarity calculations are performed on the combined tag vector matrix and the user feature vector, for example, using cosine similarity in the similarity calculation (the present invention is not limited thereto), obtaining weights for the combinations, sorting the multiple recommended combined tags based on the weights of the combinations, and providing the combined tag that best suits the user's needs. (6) The product display module 16 sets a catch, then selects products from among multiple products in the product database 123 that have the tag of the catch, arranges them based on the weight of the tags, transmits them to the user's information device 3, displays them on the user interface, and provides the product that best suits the user's needs. Preferably, the product display module 16 further sets the combination tag generated by the tag combination module 15, then selects at least one recommended product from the product database 123 that simultaneously has the combination tag, and performs a recommendation calculation for each recommended product. For example, if the combination tag is set as one catch and one sub-catch, then for example, 60% and 40% (total 100%) are used as the weight distribution ratios for the recommendation calculation. For example, if the main tagline is "body soap," the user has a purchase history of body soap, so the weight is set to 5. If the sub-tagline is "antibacterial," the user has clicked on antibacterial-related knowledge base articles, so the weight of the sub-tagline "antibacterial" is set to 2. The recommendation weights when the recommendation calculation is performed are shown below. 5 × 60% + 2 × 40% = 3.8 Each recommended product uses its recommendation weight as the basis for sorting when displaying the products. On the other hand, once the sorting of the display is complete, the product display module 16 updates and changes the above values ​​based on one or a combination of the weights of each product's tags, recommendation weights, etc. For example, if influenza has spread within the last 6 hours, the weight of the "antibacterial" tag will be significantly increased, and the product display module 16 will automatically change its release and sort the recommended products again based on the updated values, raising the ranking of "antibacterial" related products.

[0013] Referring together to Figure 2 and Figure 1, the steps of the method for carrying out the present invention will be described below. (1) Model training step S1: The real-time product recommendation system 1 performs machine learning on existing information and existing related information stored in the learning database 121 to construct a keyword model. The machine learning is one of the "Machine Learning" methods such as "Supervised Learning," "Semi-Supervised Learning," "Reinforcement Learning," "Unsupervised Learning," "Self-Supervised Learning," and "Heuristic Algorithms," or a combination thereof, but the present invention is not limited to these. (2) Real-time capture step S2: As shown in Figure 3, when a user updates a product tag and its weight, first, the capture item is set (S21). Also as shown in Figure 4, the real-time capture module 13 sets the capture type A1 and the time interval A2, where the capture type A1 refers to the type of real-time information. In this implementation, "Hot News" is used as an example, and "within 6 hours" is set as the real-time update time interval A2. Preferably, the real-time capture module 13 sets the related information type A3, and in this implementation, the "number of views" of "Hot News" (real-time information) is set as the related information. When "Real-time information collection S22" is executed, the real-time capture module 13 collects the real-time information to be captured by methods such as browser search, data capture, "Web Crawler", or a combination thereof, but the present invention is not limited thereto. Next, it checks whether the real-time information to be captured is within the time interval A2 (S23). As shown in Figure 5, the real-time capture module 13 crawls multiple real-time information (T1~T3) from the news site N, captures at least one time mark (t1~t3), and then determines whether the time marks (t1~t3) are within a set time interval (A2). For example, if the "3 hours ago" time mark t1 and the "1 hour ago" time mark t2 of this implementation both fit within a time interval (A2) of "within 6 hours," and the corresponding real-time information (T1, T2) fits the user's needs, then "Acquire real-time information S24" is executed to capture the real-time information (T1, T2) that fits within time interval A2. The number of browses (V1, V2) is used as related information, and the captured real-time information (T1, T2) is transmitted to the word extraction module 14 and stored in the learning database 121 as existing information for subsequent implementations. (3) Word extraction step S3: As shown in Figure 6, the word extraction module 14 receives the real-time information (T1, T2) from the capture mentioned above, then performs "Natural Language Processing, NLP" to analyze the real-time information (T1, T2) and extract at least one real-time word (P1, P2). For example, it extracts hot words such as those with high search frequency or topicality, or characteristic words such as those with high relevance to the product or functionality. In this implementation, "low-carbon bicycle" from the captured real-time information T1 and "Pudu" from the real-time information T2 are used as real-time words (P1, P2). As shown in Figures 7 and 8, the word extraction module 14 analyzes and processes the extracted real-time word P1 using the trained keyword model. Specifically, it generates at least one highly relevant real-time tag through information processing such as word segmentation and filtering. In this embodiment, the tag combining module 15 generates multiple highly relevant real-time tags (tag1~tag4) from "low-carbon bicycle" (P1). "Low carbon" and "bicycle" are real-time tags (tag1, tag2) that have undergone word segmentation processing, while "exercise" and "triathlon" are highly relevant real-time tags (tag3, tag4) generated by the model. After step S31, which generates the tags described above, the tag combining module 15 saves the real-time tags (tag1~tag4) to the tag database 122 and further executes step S32, which calculates the tag weights. Vector operations are performed on the real-time tags to form tag feature vectors for each tag, and vector operations are performed on user behavior data to form user feature vectors. Then, the similarity between the tag feature vectors and user feature vectors is calculated to generate tag weights. Next, the module checks whether the real-time tags (tag1~tag4) are existing tags (S33) and compares them one by one with the existing tags of the product. If the comparison result is a match, the tag weights are updated (S34), and the existing weights of the existing tags are updated as real-time weights. If, as a result of the comparison, no identical existing tags exist, step S35 is executed to attach the tags.As shown in Figure 9, in a product database 123 where multiple products are stored, the tag joining module 15 attaches real-time tags (tag1 to tag4) to multiple related candidate recommended products (R1 to R4) based on the relevant information that each real-time tag (tag1 to tag4) corresponds to, sets the tag weights (S36), and adds the real-time weights to the real-time tags (tag1 to tag4). (4) Combination tag generation step S4: As shown in Figures 10 and 11, the tag combining module 15 receives the real-time tag tag4 generated by the word extraction module 14, and then sets the real-time tag tag4 as catch M1. Catch M1 performs one or a combination thereof of artificial intelligence-related technologies such as machine learning, deep learning, and large-scale language learning using a combination tag model, but the present invention is not limited thereto. Multiple combination tags (C1 to C5) having different sub-catches (U1 to U5) are listed, and vector operations are performed on the catch M1 and sub-catches (U1 to U5) to form catch feature vectors and sub-catch feature vectors, respectively, and by forming a combination tag vector matrix, the relationship and fit of catches and sub-catches of different combination tags are calculated. Next, the combination tag vector matrix is ​​used to calculate the similarity between the combination tags (C1~C5) and the user characteristic vector F of the user behavior data. For example, the cosine similarity is calculated to obtain the weight of the combination and select the recommended combination tag group P. Based on the weight of the combination, the multiple recommended combination tags (P1~P4) are rearranged. As shown in Figure 12, when the user inputs catch M1 via the information device 3, the tag combination module 15 calculates and selects, transmits the recommended combination tags (P1~P4) to the information device 3, and displays them on the user interface 31, providing the combination tag that best suits the user's needs. (5) Automatic product release step S5: Referring to Figures 12 and 13, when the user selects a recommendation combination tag P1 using the information device 3, the recommendation combination tag P1 is transmitted to the product display module 16, where products are collected and sorted to obtain multiple candidate recommended products (R1~R4). A recommendation calculation is performed for each sorted product, and the recommendation weight is obtained for each. A second sorting is performed, and then the aforementioned three products are obtained based on the recommendation weights (R1~R3 ​​in the figure) to obtain a group of recommended products G. Referring to Figure 14, each sorted candidate recommended product is formed into recommended products (G1~G3) based on the recommendation weight, released in order, transmitted to the user's information device 3, and each recommended product is displayed on the user interface 31.

[0014] As described above, the present invention primarily captures real-time information within a time interval using a real-time capture module, analyzes the real-time information using artificial intelligence, extracts and pastes tags, collects and calculates combination tags of catch and sub-catch, combines them with user behavior data to perform vector calculations, and selects recommended products that best match the theme and user intent. Finally, by performing the above steps at regular intervals (periodically) to update and rearrange the contents of the recommended product list in real time, the problem of limitations in recommended products in conventional product recommendation systems is solved, real-time information is more accurately reflected in product recommendations, and consumer real-time needs are met. As described above, by implementing the present invention, the objective of providing a real-time product recommendation system and its implementation method that periodically captures network data, extracts words, generates relevant tags and their weights using artificial intelligence, and tags corresponding related products is reliably achieved.

[0015] Although the present invention has been described above using embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments. It will be apparent to those skilled in the art that various modifications or improvements can be made to the above embodiments. It will be clear from the claims that such modified or improved forms may also be included in the technical scope of the present invention. [Explanation of Symbols]

[0016] 1. Real-time product recommendation system 11. Central Processing Module 12 storage modules 121 Learning Database 122 Tag Database 123 Product Database 13. Real-time capture module 14 Word Extraction Module 15 Tag Combination Module 16 Product Display Module 2. Internet 3 Information equipment 31 User Interface A1 Capture Type A2 Time interval A3 Related Information Types N News Site T1~T3 Real-time Information t1~t3 Time Marks V1 Browse count V2 Browse count P1 Real-time words P2 Real-time Words tag1~tag4 Real-time tags M1 Catch U1~U5 Sub-catch C1~C5 Combination Tags F User behavior data user characteristic vector P Recommendation combination tag group P1~P4 Recommendation combination tags R1~R4 Recommended Products G Recommended Products G1~G3 Recommended Products S1 Model Training Steps S2 Real-time Capture Step S3 word extraction step S4 Combined Tag Generation Step S5 Product Automated Release Step Set up S21 capture items S22 Real-time information gathering S23 Whether it is within the time interval S24 Real-time information acquisition Generate S31 tags Calculate the weight of the S32 tag. S33 Is it an existing tag? Update the weight of the S34 tag. Attach the S35 tag. Set the tag weights for S36.

Claims

1. A real-time product recommendation system that is connected to the internet and capable of communication, It is a central processing module, A storage module including a learning database and a product database is communicatively connected to the central processing module, the learning database stores keyword models, and the product database stores multiple "product information" entries. A real-time capture module that is communicatively connected to the central processing module and captures real-time information via the internet during a time interval, A word extraction module is communicatively connected to the central processing module, extracts at least one real-time word from the real-time information, generates at least one real-time tag from the real-time word using the keyword model in the learning database, transmits the real-time tag to a tag combining module communicatively connected to the central processing module, generates a combined tag consisting of a catch and a sub-catch, and attaches the real-time tag to at least one "product information" in the product database based on the tag weights. A real-time product recommendation system comprising: a product display module that is communicably connected to the central processing module, the product display module selects a plurality of product information in the product database using the combination tag, performs a recommendation calculation to generate recommendation weights, and then generates at least one recommended product by the product display module.

2. The real-time product recommendation system according to claim 1, characterized in that the real-time capture module captures related information of the real-time information, the related information is used to calculate the weight of the tag, and the tag coupling module updates the real-time tag based on the related information.

3. The real-time product recommendation system according to claim 2, characterized in that the aforementioned related information is one or a combination thereof of the number of browses, the number of reach, the number of likes, the number of comments, the number of shares, the number of purchases, and the browse record.

4. The real-time information includes user behavior data, and when the word extraction module is executed, it generates a tag feature vector and a user feature vector by performing vector operations on the real-time tag and the user behavior data, respectively, calculates the similarity between the tag feature vector and the user feature vector, and then generates the weight of the tag, as described in claim 1.

5. The real-time product recommendation system according to claim 4, characterized in that the tag merging module performs vector operations on the combination tags to generate a combination tag vector matrix, calculates the similarity between the combination tag vector matrix and the user feature vector, generates weights for the combinations, and the tag merging module generates a plurality of recommendation combination tags based on the weights for the combinations.

6. A real-time capture step in which a real-time capture module captures real-time information over the internet during a time interval, A word extraction step in which a word extraction module extracts at least one real-time word from the real-time information, generates at least one real-time tag from the real-time word using a keyword model, and attaches the real-time tag to at least one related product in the product database of the storage module based on the tag weights, The combination tag generation step involves the tag combining module receiving the real-time tags generated by the word extraction module and generating multiple combination tags using a combination tag model, A method for implementing a real-time product recommendation system, characterized by including a product display module that selects a plurality of products stored in the product database using the combination tags, performs a recommendation calculation to obtain recommendation weights for each, and the product display module generates at least one recommended product based on the recommendation weights in an automated product release step.

7. The method for implementing the real-time product recommendation system according to claim 6, characterized in that, in the real-time capture step, the real-time capture module captures relevant information of the real-time information for calculating the weight of the tag.

8. The method for implementing the real-time product recommendation system according to claim 7, characterized in that, in the step of attaching the real-time tag, the tag combining module identifies those products from the related products that have an existing tag as the real-time tag, and updates the weight of the existing tag based on the related information.

9. The method for implementing the real-time product recommendation system according to claim 7, characterized in that, in the word extraction step, the word extraction module performs vector operations on the real-time tag and user behavior data to generate a tag feature vector and a user feature vector, respectively, and then calculates the similarity between the tag feature vector and the user feature vector to generate the weight of the tag.

10. The method for implementing the real-time product recommendation system according to claim 9, wherein in the combination tag generation step, the tag combining module generates a combination tag vector matrix by performing vector operations on the plurality of combination tags, then generates combination weights by performing a similarity calculation between the combination tag vector matrix and the user feature vector, and the tag combining module generates a plurality of recommendation combination tags based on the combination weights.

Images