An advertisement item recommendation method and system

Abstract

The application relates to an advertisement commodity recommendation method and system, which comprises the following steps: obtaining a commodity click sequence of a user from a user log in chronological order; obtaining commodity information in the commodity click sequence and sequentially connecting the commodity information together to form user click content according to the commodity click sequence; generating a user vector according to the user click content; calculating the distance between the user vector of each user and the user vectors of other users, obtaining a preset number of user vectors with the minimum distance, and thus obtaining a similar user set of each user; and when a condition of recommending commodities to a target user is met, obtaining advertisement commodities clicked by similar users of the target user to form an advertisement commodity recommendation set, and recommending the advertisement commodities in the advertisement commodity recommendation set to the target user. The application improves the recommendation effect of the advertisement commodities and improves the coverage of the advertisement commodities.

Description

Technical Field

[0001] This invention relates to the field of internet advertising, and in particular to a method and system for recommending advertised products. Background Technology

[0002] In the field of internet advertising, ad recommendation systems need to identify target users and then recall suitable ad products for them. Ad placement typically involves setting a target audience. A common process is for advertisers to provide seed users, and the ad distribution platform uses machine learning to predict similar users, thereby expanding the ad audience. When a user logs on the platform, the ad recommendation system can identify the target user and recall ads for that user as the target audience. Common recall methods include collaborative filtering, content-based recall using natural language processing, and vector recall. Generally, a user may have multiple ads as the target audience, and the recommendation system ranks the recalled ads based on factors such as ad payment and pushes them to the user in sequence. However, for secondhand platforms, where ad products are sparse, there are problems with insufficient ad product coverage and inaccurate recommendations when recommending ad products to users. Summary of the Invention

[0003] In view of the technical problems existing in the prior art, the present invention proposes an advertising product recommendation method and system to improve the recommendation effect of advertising products and increase the coverage of advertising products.

[0004] To address the aforementioned technical problems, according to one aspect of the present invention, an advertising product recommendation method is provided, comprising: obtaining a user's product click sequence from a user log in chronological order; obtaining product information from the product click sequence and sequentially concatenating the product information according to the product click sequence to form user click content; generating user vectors based on the user click content; calculating the distance between each user's user vector and the user vectors of other users to obtain a preset number of user vectors with the smallest distance, thereby obtaining a set of similar users for each user; and when the conditions for recommending products to a target user are met, obtaining advertising products clicked by similar users of the target user to form an advertising product recommendation set, and recommending advertising products from the advertising product recommendation set to the target user.

[0005] According to another aspect of the present invention, an advertising product recommendation system is provided, comprising a product click sequence construction module, a user click content construction module, a user vector acquisition module, a similar user acquisition module, and a recommendation module. The product click sequence construction module acquires a user's product click sequence from a user log in chronological order. The user click content construction module is connected to the product click sequence construction module and configured to acquire product information from the product click sequence, and sequentially connects the product information according to the product click sequence to form user click content. The user vector acquisition module is connected to the user click content construction module and configured to obtain corresponding user vectors based on the user click content. The similar user acquisition module is connected to the user vector acquisition module and configured to calculate the distance between each user's user vector and the user vectors of multiple other users, and to form a set of similar users for a given user by selecting a preset number of users with the smallest vector distance. The recommendation module is connected to the similar user acquisition module, and when the conditions for recommending advertising products to a target user are met, it acquires advertising products clicked by similar users of the target user to form an advertising product recommendation set, and recommends advertising products from the advertising product recommendation set to the target user.

[0006] This invention projects the similarity of clicked products onto the similarity of users, thereby expanding the target user's similar user group and increasing the coverage of advertised products. This solves the problem of insufficient coverage of advertised products during recommendation due to sparse product offerings. Because of the correlation between user similarity and clicked product similarity, the accuracy of recommending advertisements to target users is improved, effectively increasing the click-through rate of advertised products and boosting the revenue of both the platform and advertisers. Attached Figure Description

[0007] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0008] Figure 1 This is a flowchart of an advertising product recommendation method according to an embodiment of the present invention;

[0009] Figure 2 This is a flowchart for determining a set of similar users for a target user according to an embodiment of the present invention;

[0010] Figure 3 This is a schematic diagram of user click content generation according to an embodiment of the present invention;

[0011] Figure 4 This is a block diagram illustrating the principle of an advertising product recommendation system according to an embodiment of the present invention;

[0012] Figure 5This is a block diagram of a user vector acquisition module according to an embodiment of the present invention; and

[0013] Figure 6 This is a block diagram of the recommended module provided according to an embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0015] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.

[0016] Figure 1 This is a flowchart of an advertising product recommendation method according to an embodiment of the present invention. The advertising product recommendation method provided by the present invention includes the following steps:

[0017] Step S1: Determine a set of similar users for the target user.

[0018] Step S2: Obtain a set of recommended advertising products that have been clicked by similar users.

[0019] Step S3: Sort the products in the recommended product set. For example, sort them according to the advertising delivery method and payment method, using different metrics, such as CTR (Click Through Rate) or eCPM (Effective Cost per Mille), or do not sort them.

[0020] Step S4: Determine if the recommendation criteria are met. If they are met, in Step S5, determine the target recommended advertising product and recommend it to the target user. If not, wait and return to Step S4. The number of advertising products recommended varies depending on the target user's scenario. For example, when a target user logs into the app, in the homepage product display scenario, the number of products that can be displayed is determined based on the user's screen size, and that number of advertising products is selected from the recommended product set and displayed in a preset position on the homepage. When a user swipes down on the screen, the number of additional products to be displayed is determined based on the swipe distance, and that number of advertising products is selected from the recommended product set and displayed in the required preset position. This can also be applied to other scenarios where advertising products need to be displayed, which will not be elaborated here.

[0021] In step S1, as follows Figure 2 One embodiment shown illustrates the process of determining a set of similar users for a target user, as follows:

[0022] Step S11: Query the user log to determine the products clicked by the user in the past, forming a product click sequence. Platforms using this invention, such as second-hand e-commerce platforms, record user behavior to form user logs. These logs record user actions on the platform, such as searching for products, clicking, viewing products, placing orders, and posting comments after purchase. For each action, corresponding information is recorded, such as the time and keywords of the search, the time of clicking on a product, and the product's identifier. By querying the user logs, the products clicked by the user within a preset time period can be obtained, and a product click sequence can be obtained in chronological order, such as... Figure 3 Item 1 and Item 2 are shown. The number of items in this sequence can be determined by a preset time period or a preset quantity value. For example, the item click sequence can be obtained by arranging all the items clicked by the user in the previous week in chronological order, or the item click sequence can be generated from the latest 10 or 20 items in the user log.

[0023] Step S12: Retrieve product information from the product click sequence in the platform's product information database or product information table. The product information includes the product title and description. For example... Figure 3the title A1 of the commodity 1 and the description information A2, and the title B1 of the commodity 2 and the description information B2 obtained in. The A1 and B1 described herein are usually composed of several short words, and A2 and B2 are usually a piece of text. For example, the title A1 of the commodity item1 is "95% new iphone X", and the description information A2 is "Personal use iphone X, newly bought and not old, with intact screen, contact me soon please". The title B1 of the commodity item 2 is "Lightweight men's sportswear", and the description information B2 is "XL size lightweight men's top sportswear, windproof and breathable, loose version".

[0024] In step S13, the commodity information is connected in sequence according to the commodity click sequence to form user click content. For example Figure 3 A1 A2 B1 B2 are obtained in, for example, "95% new iphone X Personal use iphone X, newly bought and not old, with intact screen, contact me soon please Lightweight men's sportswear XL size lightweight men's top sportswear, windproof and breathable, loose version".

[0025] In step S14, a user vector is generated according to the user click content. The user click content is a string of text. To convert this text into a vector, in one embodiment, semantic analysis is first performed on the user click content, and word segmentation is performed according to the semantics. Any word segmentation algorithm, such as the CRF algorithm, may be used to perform word segmentation on the user click content. For example, after word segmentation is performed on the user click content "95% new iphone X Personal use iphone X, newly bought and not old, with intact screen, contact me soon please Lightweight men's sportswear XL size lightweight men's top sportswear, windproof and breathable, loose version", "95% new iphone X iphone X personal use, newly bought and not old, with intact screen, contact me soon please lightweight men's sportswear XL size lightweight men's top sportswear, windproof and breathable, loose version" is obtained, so as to obtain a click word sequence. Then the click word sequence is filtered, for example, punctuation marks therein are filtered out, and some words or phrases that have no specific semantics in the second-hand scenario are filtered out with reference to a general filtered word list and a second-hand scenario-specific filtered word list, such as the aforementioned "contact me soon please". The general filtered word list records meaningless modal particles, auxiliary words, pronouns and other words in the Chinese context, such as "de (of)", "ya (ah)", "de (get)", "wo (I)", "ni (you)", etc. The second-hand scenario-specific filtered word list records words and sentences that appear frequently on second-hand platforms and have no corresponding meaning to the item title itself, such as "second-hand", "please contact me", "platform transaction", etc. The filtered click word sequence is, for example, "95% new iphone x personal use newly bought not old intact screen lightweight men's sportswear XL size lightweight men's top sportswear windproof breathable loose version".

[0026] Step S15: Obtain the word vector of each word segment constituting the clicked word sequence. If a word vector library is stored in the system, the word vector of each word segment can be read from the library; if the word vector library is not available, a natural language semantic model can be trained using a certain vocabulary, and then the word vector of each word segment in the clicked word sequence can be obtained from the natural language semantic model. The natural language semantic model is, for example, the word2vec model.

[0027] Step S16: Calculate the average of all word vectors to obtain the vector of the clicked word sequence, and use the vector of the clicked word sequence as the user vector. For example, the word vector of "95 new" is [0.1, 0.5, -0.3], and the word vector of "iphonex" is [0.9, 0.2, -0.1]. After average pooling, the average of these two word vectors can be obtained as [0.5, 0.35, -0.2].

[0028] Step S17: Calculate the distance between each user's user vector and the user vectors of other users. Steps S11 to S16 yield the user vectors of all users on the platform to which this invention is applied, such as a second-hand e-commerce platform. The calculated vector distances are then sorted, and the top N users with the smallest distances are selected as the similar users of that user, thus obtaining the similar user set for each user.

[0029] This invention addresses the problem of sparse advertising products and insufficient coverage during recommendations by expanding the user base to increase the coverage of advertised products. Since the expanded user base consists of users similar to the target users, the accuracy of recommendations is also improved, thereby effectively increasing the click-through rate for advertisers and boosting revenue for both the platform and advertisers. In some embodiments, the advertised products clicked by a user also include one or more items deleted by the user, and these deleted items are placed at the beginning of the user's click sequence. This also improves the accuracy of advertised product recommendations.

[0030] According to another aspect of the present invention, the present invention also provides an advertising product recommendation system, such as... Figure 4As shown, the advertising product recommendation system includes a product click sequence construction module 1, a user click content construction module 2, a user vector acquisition module 3, a similar user acquisition module 4, and a recommendation module 5. The product click sequence construction module 1 is used to obtain the user's product click sequence from the user log in chronological order. It can obtain the products clicked by the user from the user log within a preset time period to form the product click sequence; or it can obtain a preset number of products clicked by the user from the user log in chronological order as the product click sequence. The user click content construction module 2 is connected to the product click sequence construction module 1 and is used to obtain product information from the product click sequence and connect the product information sequentially according to the product click sequence to form the user click content. The user vector acquisition module 3 is connected to the user click content construction module 2 and is used to generate user vectors based on the user click content. In one embodiment, such as... Figure 5As shown, the user vector acquisition module 3 includes a word segmentation unit 31, a filtering unit 32, a word vector acquisition unit 33, and a word sequence vector calculation unit 34. The word segmentation unit 31 segments the user-clicked content using a semantic analysis algorithm to obtain a clicked word sequence. This segmentation can employ algorithms such as CRF or dictionary comparison, which will not be elaborated further here. The filtering unit 32 is connected to the word segmentation unit 31 and is used to filter the segmented user-clicked content. For example, it removes punctuation marks and removes words without specific meanings in the secondhand e-commerce market by querying a general filtering vocabulary and a secondhand scenario-specific filtering vocabulary. The word vector acquisition unit 33 is connected to the filtering unit 32 and is used to obtain the word vector for each segmented word in the clicked word sequence. When the system can access the word vector library 6, it obtains the word vector for each segmented word in the clicked word sequence from the word vector library 6. If the word vector library 6 does not contain the word vector, a natural language semantic model, such as the word2vec model, is used for training using a vocabulary of commonly used words in the secondhand e-commerce platform. The click word sequence obtained in this embodiment is then input into the natural language semantic model to obtain the word vector for each segment. In this embodiment, the word vectors are obtained using a vocabulary of commonly used words from the second-hand e-commerce platform as the corpus, such as "80% new," "90% new," "wear and tear," and "scratches." Therefore, to improve the efficiency of each recommendation and reduce computation, one embodiment also includes a word vector library 6 and a semantic model module 7. The semantic model module 7 uses the vocabulary accumulated on the platform to form a vocabulary list, and a portion of this list is used as a training set to train the word2vec model. The trained word2vec model is used to obtain the word vectors from the vocabulary list, which are then stored and used to form the word vector library 6 for later use. Whenever a word vector is needed, the word vector library 6 is queried first. If the word vector for an unwanted segment is obtained from the word vector library 6, the word2vec model is called to obtain the word vector for that segment, and the newly obtained word vector is stored in the word vector library 6, adding the word to the vocabulary list. This satisfies the recommendation requirements during business implementation and also expands the vocabulary list and word vector library. The word sequence vector calculation unit 34 is connected to the word vector acquisition unit 33 and is used to calculate the average of the word vectors of all word segments based on the word vector of each segment to obtain the vector of the word sequence, and use the vector of the word sequence as the user vector. In one embodiment, the word sequence vector calculation unit 34 performs an Average Pooling operation on all word vectors in the clicked word sequence example to obtain the vector of the word sequence. This word sequence vector can then be used as the user vector. Since the user vector comes from the information of the products it has clicked, the user vector actually represents the characteristics of the product information. If the distance between two user vectors is small, it indicates that the product information of the two is also similar.

[0031] The similar user acquisition module 4 is connected to the user vector acquisition module 3. It is configured to calculate the distance between each user's user vector and the user vectors of multiple other users, and to form a set of similar users for each user by selecting a predetermined number of users with the smallest vector distances. Various methods can be used to calculate vector distances. For example, Faiss (Facebook AI Similarity Search) can be used to perform nearest neighbor retrieval based on a user vector within the user vector set, thus obtaining the user vectors closest to that user vector. Nearest neighbor retrieval involves calculating the vector distance between a user vector and all other user vectors, and selecting the top N user vectors with the smallest distances. Faiss is a tool to accelerate this calculation process; other tools can also be used, such as Milvus (an open-source vector similarity search engine) or Annoy (Spotify's open-source approximate nearest neighbor query library). After processing by the similar user acquisition module 4, a set of similar users can be obtained for each user on the second-hand e-commerce platform. Based on the advertised products clicked by each similar user, a set of recommended advertised products can be obtained for each user.

[0032] Recommendation module 5 is connected to the similar user acquisition module 4. When the conditions for recommending products to the target user are met, it acquires the target user's set of recommended advertised products and recommends advertised products from the set to the target user. In one embodiment, such as... Figure 6As shown, the recommendation module 5 includes a triggering unit 51, an advertising product acquisition unit 52, and a recommendation unit 53. The triggering unit 51 issues a recommendation notification when conditions are met to recommend advertising products to the target user. For example, it issues a notification when a user is detected logging into the app's homepage, or when a user is detected on a page with advertising space. The advertising product acquisition unit 52 is connected to the triggering unit 51 and acquires the target user's advertising product recommendation set upon receiving the recommendation notification. The recommendation unit 53 is connected to the advertising product acquisition unit 52 and is used to push advertising products from the recommended advertising product set to the target user based on their current scenario. Specifically, the recommendation unit 53 acquires a corresponding number of advertising products from the advertising product push set based on the current scenario. For example, when the user is on the app's homepage, it acquires a corresponding number of advertising products based on the number of products that can be displayed on the homepage; or, when the user swipes the screen, it acquires the number of advertising products corresponding to that swipe distance. Then, it displays the advertising products in the corresponding positions on the page. In one embodiment, after obtaining the corresponding number of advertising products, the advertising products can be sorted, for example, by using different sorting indicators according to the payment method of the advertisement, such as sorting by CTR or by eCPM, to obtain an ordered set of advertising products; the recommendation unit 53 sequentially obtains the number of advertising products corresponding to the scene from the set of advertising products and displays them in the corresponding positions.

[0033] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.

Claims

1. A method for recommending advertised products, characterized in that, include: Retrieve the user's product click sequence from the user logs in chronological order; The product information in the product click sequence is obtained, including product title and product description information, and the product information after connecting the product title and product description information is connected in order according to the product click sequence to form the user click content; Generate a user vector based on the user's click content; Calculate the distance between each user's user vector and the user vectors of other users, obtain the preset number of user vectors with the smallest distance, and thus obtain the set of similar users for each user; as well as When the conditions for recommending products to the target user are met, the advertising products clicked by similar users of the target user are obtained to form an advertising product recommendation set, and the advertising products in the advertising product recommendation set are recommended to the target user. The step of generating a user vector based on the user's click content specifically includes: The user-clicked content is semantically analyzed and filtered to obtain a click word sequence composed of multiple word segments; Obtain the word vector for each segment that constitutes the clicked word sequence; and Calculate the average of all word vectors to obtain the vector of the clicked word sequence, and use the vector of the clicked word sequence as the user vector.

2. The method according to claim 1, characterized in that, Further includes: Obtain the user's product click sequence from the user log within a preset time period; or obtain a preset number of products clicked by the user from the user log in chronological order as the product click sequence.

3. The method according to claim 1, characterized in that, The steps of performing semantic analysis and filtering on the user-clicked content further include: Perform semantic analysis on the content clicked by the user to obtain multiple word segments; Query general filter terms and secondhand scenario-specific filter terms; and Remove word segments that appear in the general filter thesaurus or the secondary scene-specific filter thesaurus.

4. The method according to claim 1, characterized in that, Further includes: The word vectors of each segment that constitute the clicked word sequence can be obtained through the word2vec model, or the word vectors of each segment can be read from a word vector library.

5. The method according to claim 4, characterized in that, Further includes: The system queries a pre-defined vocabulary list. In response to finding a word segment in the clicked word sequence in the vocabulary list, the system retrieves the corresponding word vector from the word vector library. When a word segment cannot be found in the clicked word sequence in the vocabulary, the word vector of the word segment is obtained through the word2vec model, and the word segment is recorded in the vocabulary and the corresponding word vector is recorded in the word vector library.

6. An advertising product recommendation system, characterized in that, include: The product click sequence construction module is configured to retrieve the user's product click sequence from the user log in chronological order; The user click content construction module is connected to the product click sequence construction module. It is configured to obtain product information in the product click sequence. The product information includes product title and product description information. The product information is then connected sequentially according to the product click sequence to form the user click content. The user vector acquisition module is connected to the user click content construction module and is configured to obtain the corresponding user vector based on the user click content. A similar user acquisition module, which is connected to the user vector acquisition module, is configured to calculate the distance between each user's user vector and the user vectors of multiple other users, and to form a set of similar users for a user by selecting a preset number of users with the smallest vector distance. as well as The recommendation module, which is connected to the similar user acquisition module, when the conditions for recommending advertising products to the target user are met, acquires the advertising products clicked by similar users of the target user to form an advertising product recommendation set, and recommends advertising products in the advertising product recommendation set to the target user. The user vector acquisition module includes: The word segmentation unit is configured to perform semantic analysis on the user's clicked content, and to divide the user's clicked content into multiple words to obtain the initial clicked word sequence; A filtering unit, connected to the word segmentation unit, is configured to remove word segments that appear in the general filter word list and the secondary scene-specific filter word list from the initial click word sequence to obtain the click word sequence; A word vector acquisition unit, configured to obtain the word vector of each word segment in the clicked word sequence; and A word sequence vector calculation unit, connected to the word vector acquisition unit, is configured to obtain the vector of the clicked word sequence by calculating the average value of the word vectors of all segments in the clicked word sequence, and use the vector of the clicked word sequence as the user vector.

7. The system according to claim 6, characterized in that, The product click sequence construction module is configured to obtain the user's product click sequence from the user log within a preset time period; or to obtain a preset number of products clicked by the user from the user log in chronological order as the product click sequence.

8. The system according to claim 6, characterized in that, It also includes a word vector library and a natural language semantic model module. The word vector library stores word vectors of words in a pre-set vocabulary. The natural language semantic model module trains a natural language semantic model using training corpus to generate corresponding word vectors based on a given vocabulary.

9. The system according to claim 6, characterized in that, The recommendation module includes: The triggering unit is configured to issue a recommendation notification when conditions are met to recommend advertised products to the target user; An advertising product acquisition unit, connected to the triggering unit, is configured to acquire advertising products that have been clicked by similar users of the target user, forming an advertising product recommendation set; and The recommendation unit, which is connected to the advertising product acquisition unit, is configured to recommend advertising products from the set to the target user based on the user's current location.

Images