Content recommendation method and system based on user tag matrix and storage medium
By building a static and dynamic tag matrix of user identity and behavior, combining content recommendation and collaborative filtering algorithms, we generate and optimize the recommended content queue, and solve the problem of content homogeneity and computing complexity, and achieve personalized and efficient content recommendation.
Patent Information
- Application Number
- CN202510298020.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-13
- Publication Date
- 2025-08-01
AI Technical Summary
The existing recommendation algorithms have serious content homogeneity in content recommendations and it is difficult to maintain the relevance and accuracy of recommendations, especially when the number of users increases, the calculation overhead is too large.
A static and dynamic tag matrix based on user identity and behavior is constructed, combined with content recommendation algorithm and collaborative filtering algorithm, user identity characteristics and dynamic tag matrix are extracted through static tag matrix, behavior preference characteristics are extracted, initial recommended content queue is generated, and the preferred content of similar user groups is supplemented through collaborative filtering algorithm, and the weight of recommended content is calculated for sorting.
It improves the accuracy and diversity of recommended content, reduces calculation complexity, improves processing efficiency and stability, ensures that recommended content is sorted by user interests and preferences, and improves information relevance and attractiveness.
Smart Images
Figure CN120407816A_ABST
Abstract
Description
Technical Field
[0001] The present invention specifically relates to a content recommendation method, system and storage medium based on a user tag matrix, belonging to the technical field of recommendation algorithms. Background Art
[0002] The recommendation algorithm is a commonly used technology in recommendation systems. It combines technologies such as big data, machine learning and mathematical algorithms to perform data analysis on information such as user behavior preferences, so as to provide push content that meets user needs. An effective recommendation algorithm can discover the potential preferences of users, accurately perform personalized content recommendation, improve the quality of content production, and enhance the user experience of the system; moreover, for developers, it is beneficial to accelerate the information supply efficiency, enhance the activity of the system, and increase the access volume of the system.
[0003] Common recommendation algorithms are mainly divided into content recommendation algorithms and collaborative filtering algorithms. The content recommendation algorithm is a recommendation algorithm model constructed based on data such as relevant information of the target, user information, and user behavior characteristics; the collaborative filtering algorithm is based on the similarity between users, adhering to the principle that user groups with similar behavior preferences have similar preferences, and recommends information to users. However, it is obvious that the content recommendation algorithm has serious homogenization in content recommendation, the recommended content lacks novelty, and at the same time, as the content is continuously updated, it is difficult for the algorithm to maintain the relevance and accuracy of the recommendation; while the collaborative filtering algorithm needs to classify user groups, and when the number of users grows too fast, it is easy to cause excessive data calculation overhead. It can be seen that the problems existing in the current recommendation algorithm need to be overcome. Summary of the Invention
[0004] The object of the present invention is to propose a content recommendation scheme based on static and dynamic user tag matrices, and accurately push content information that meets user preferences according to user identity information and conditional restrictions, so as to improve the accuracy of information push and the data processing efficiency of the system.
[0005] The technical solutions adopted by the present invention to achieve the above object are as follows:
[0006] A content recommendation method based on a user tag matrix, comprising the following steps:
[0007] 1) Construct a static tag matrix based on user identity information, and construct a dynamic tag matrix based on user operation behavior information;
[0008] 2) Extract the behavior preference characteristics of the user from the dynamic tag matrix, calculate the preference degree of the user for the recommended content using the content recommendation algorithm, and generate an initial recommended content queue according to the preference degree;
[0009] 3) Extract the identity features of the user from the static label matrix, use the collaborative filtering algorithm to find the group of similar users of the target user, supplement the content preferred by the group of similar users to the initial recommended content queue, and optimize the recommended content;
[0010] 4) For the optimized recommended content, calculate the preference degree of the target user group for the recommended content as the weight of the recommended content, sort the recommended content according to the weight, and recommend it to the target user in order.
[0011] Further, each row of the static label matrix in step 1) represents the identity information of a user, which consists of multiple identity features; each row of the dynamic label matrix represents the operation behavior information of a user, which consists of multiple behavior preference features.
[0012] Further, the method for calculating the preference degree of the user for the recommended content using the content recommendation algorithm in step 2) includes: extracting the keyword features of the system articles, calculating the cosine similarity between the user's behavior preference features and the keyword features of the system articles, and using the cosine similarity as the preference degree of the user for the recommended content.
[0013] Further, the method for generating the initial recommended content queue according to the preference degree in step 2) includes: taking the system articles with the preference degree exceeding a preset threshold as the recommended content to generate the initial recommended content queue.
[0014] Further, the method for using the collaborative filtering algorithm to find the group of similar users of the target user in step 3) includes: calculating the similarity between the target user and other users through the Pearson correlation coefficient according to the user's identity feature information, and taking the other users with the similarity greater than a preset threshold as the group of similar users of the target user.
[0015] Further, the content preferred by the group of similar users supplemented to the initial recommended content queue in step 3) refers to the system articles preferred by the group of similar users calculated using the content recommendation algorithm, and the article content obtained after removing the articles that the target user has already viewed.
[0016] Further, the steps for calculating the preference degree of the target user group for the recommended content in step 4) include:
[0017] Calculate the similarity sim(u, u i ) between the target user u and other users u i , where the other user u i refers to other users similar to the similar users of the target user u;
[0018] Calculate the preference degree score(u i , s) of other users u i for the recommended content s according to the content recommendation algorithm;
[0019] Multiply sim(u, u i ) by score(u i , s), and sum over all other users u i to obtain the preference degree sim(u, s) of the target user group for the recommended content.
[0020] A content recommendation system based on a user tag matrix, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.
[0021] A computer-readable storage medium stores a computer program, and when the computer program is executed, the steps of the above method are implemented.
[0022] The beneficial effects achieved by the present invention are as follows:
[0023] 1. By constructing static and dynamic tag matrices based on a small amount of user information data, dividing the data by dimension, and applying different recommendation algorithms, the present invention enhances the feature expression ability of the data.
[0024] 2. By hierarchically applying traditional recommendation algorithms instead of simply combining and calling single recommendation algorithms, and combining user identity and behavior data, the present invention realizes the accurate push of recommended content and improves the accuracy of the recommended content.
[0025] 3. By combining the collaborative filtering algorithm with the content recommendation algorithm, the present invention makes up for the problem of content homogenization caused by single recommendation algorithms, makes the recommended content more diverse and personalized, reduces the computational complexity, and reduces the problems of data sparsity and overfitting, improving the processing efficiency and stability.
[0026] 4. The present invention calculates the weights of the recommended content according to the user behavior preferences, ensuring that the recommended content is sorted according to the user interests and preferences, thereby improving the relevance and attractiveness of the pushed information. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 is a flowchart of the content recommendation method based on the user tag matrix in the embodiment.
[0028] Figure 2 is an example diagram of the user static tag matrix.
[0029] Figure 3 is an example diagram of the user dynamic tag matrix.
[0030] Figure 4 is an example diagram of the keyword features of the articles with higher user click-through rates.
[0031] Figure 5It is an example diagram for comparing the features of system articles and user-preferred content. Detailed implementation manners
[0032] To make the technical features, advantages or technical effects in the above technical solutions of the present invention more obvious and understandable, the following will be described in detail with reference to the accompanying drawings.
[0033] An embodiment of the present invention provides a content recommendation method based on a user tag matrix, and its processing flow is as Figure 1 shown, specifically including the following steps:
[0034] The first step: construct a static tag matrix and a dynamic tag matrix
[0035] According to the identity information of the user, construct a static tag matrix. Each row in the static tag matrix is a static tag vector X, which is composed of multiple identity features x i and represents the identity feature information of the user. Each identity feature is respectively the value (not limited to this) of the attribute information such as the gender, age, occupation, identity identifier, etc. of the user, and can be expressed as X=(x1, x2,..., x j ). These attribute information have strong stability and do not change easily. Suppose there are i users and n identity features (a1, a2,..., a n ), then a static tag matrix of R (i*n) dimensions can be constructed, as Figure 2 shown. For example, the identities of two users are a 30-year-old female Internet R & D staff and a 40-year-old male office worker in an association. By defining the attributes (gender, age, occupation, identity) and taking the values of the attribute information, that is, for gender: female takes the value 1, male takes the value 0; for age: the value is determined according to the interval, 20 - 30 years old takes the value 2, 30 - 40 years old takes the value 3; for occupation: Internet takes the value 3, association takes the value 2; for identity: R & D staff takes the value 2, office clerk takes the value 1, then the static tag vectors X1, X2 in the static tag matrix can be (1, 0, 2, 3, 2) and (0, 0, 3, 2, 1) respectively.
[0036] According to the operation behavior information of the user, construct a dynamic tag matrix (see Figure 3 ). Each row in the dynamic tag matrix is a dynamic tag vector Y, which is composed of multiple behavior preference features y i and represents the behavior preference of the user for the subject matter. The subject matter is the category of articles browsed, such as tourism, architecture, finance, philosophy, etc. (not limited to this), and the click volume on each category of articles is used as the y i value, or it can also be the article keywords, and the statistical quantity of the keywords in the article is used as the y i value, such as rising, stocks, investment, etc. (not limited to this), and can be expressed as Y=(y1, y2,..., y j), these behavioral data will fluctuate over time, interests and other factors. Suppose there are i users and n preference features (q1, q2, …, q n ), then a Q (i*n) -dimensional dynamic label matrix can be constructed, as shown in Figure 3 . For example, among all the articles read by a user, the reading volume of keywords such as tourism and architecture is counted. If the frequencies of appearance in the overall articles read are 20, 10, 30, then the dynamic label vector Y of this user is (2, 1, 3).
[0037] Step 2: Generate an initial recommended content queue based on the content recommendation algorithm
[0038] First, based on the user's dynamic label matrix, extract the user's behavioral preference features. For example, if a user has a large number of clicks on a certain article, then extract the keywords from the title and high-frequency words of this article. If the extracted keyword is "finance", then for this user, "finance" articles can be used as a reference for subsequent use of the content recommendation algorithm.
[0039] Then, extract the keyword features of the articles in the system for the calculation of the subsequent content recommendation algorithm. The extracted keyword features are as shown in the example of Figure 4 .
[0040] Finally, use the content recommendation algorithm (i.e., the nearest neighbor classification algorithm). According to the user's behavioral preference features and the keyword features of the system articles, calculate the cosine similarity as the user's preference degree for the recommended content, and add the article content with a higher preference degree (such as exceeding a preset threshold) to the initial recommended content queue. The calculation formula of the cosine similarity is as follows:
[0041]
[0042] where z i represents the user's behavioral preference features, and z i represents the keyword features of the system articles.
[0043] Figure 5 is an example diagram of the feature comparison between the system articles and the user's preferred content. If the system articles have as many features as possible in common with the user, then add the corresponding articles to the initial recommended content queue.
[0044] Step 3: Optimize the recommended content queue based on the collaborative filtering algorithm
[0045] Since the articles in the initial recommended content queue are generated according to the user's preference data, they have strong homogeneity. To solve this problem, use the collaborative filtering algorithm to supplement the data of the recommended content information.
[0046] First, find the user group similar to the target user. Specifically, according to the identity feature information in the static tag matrix of the user, calculate the similarity between the target user and other users through the Pearson correlation coefficient. If the similarity is greater than a preset threshold, determine the user group with similar characteristics to the target user. The calculation formula of the Pearson correlation coefficient is as follows:
[0047]
[0048] Among them, r represents the Pearson correlation coefficient, x i represents the identity feature of the user, represents the average value of the identity features of the user, y i represents the behavioral preference feature of the user, represents the average value of the behavioral preference features of the user.
[0049] Then, supplement the content information that the similar user group prefers and the target user has not seen (which can be calculated according to the content recommendation algorithm) into the initial recommended content queue for optimization, as shown in the example in Table 1.
[0050] Table 1 Example of supplementary data for the recommended content queue
[0051]
[0052] Step 4: Sort the recommended content
[0053] For the optimized recommended content queue, calculate the preference degree of the target user group for the recommended content in the queue and use it as the weight of the recommended content. Sort the recommended content according to the weight size, and recommend the recommended content to the target user in the order of this sorting.
[0054] The calculation formula for the preference degree of the target user group for the recommended content in the queue is as follows:
[0055]
[0056]
[0056] i ,s) represents the preference degree of other users u i similar to the similar users of the target user u for the recommended content s (which can be calculated according to the content recommendation algorithm), sim(u,u i ) represents the similarity between each target user u and other users u i , and U represents the set of other users similar to the similar users of the target user.
[0057] Table 2 shows an example of the final recommended content queue, which sorts the recommended content according to the weight (i.e., preference degree sim) size.
[0058] Recommended Content Queue Sorted by Weight in Table 2
[0059] Article title sim "Financial crisis" 31.24 "Economic Daily" 28.95 "Natural science" 25.2 "electromagnetism" 19.77 "Tourism development" 17.79 "travel" 16.75 "Natural science" 12.21 "Work report" 2.07 "fitness" 1.23
[0060] Although the present invention has been disclosed above by way of examples, it is not intended to limit the present invention. Any appropriate modifications or equivalent replacements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope of the present invention. The protection scope of the present invention shall be subject to what is defined by the claims.
Claims
1. A content recommendation method based on a user tag matrix, characterized in that, The following steps are involved: 1) Build a static label matrix based on the user's identity information and a dynamic label matrix based on the user's operation behavior information; 2) Extracting user behavior preference features from the dynamic tag matrix, using a content recommendation algorithm to calculate the user's preference for the recommended content, and generating an initial recommended content queue based on the preference; 3) Extracting user identity features from the static tag matrix, using collaborative filtering algorithms to find similar user groups to the target user, adding content preferred by similar user groups to the initial recommendation queue to optimize the recommended content; 4) For the optimized recommended content, calculate the target user group's preference for the recommended content and use it as the weight of the recommended content. Then, sort the recommended content according to the weight and recommend it to the target users in order.
2. The method according to claim 1, characterized in that, In step 1), each row of the static label matrix represents the identity information of a user, which is composed of multiple identity features; each row of the dynamic label matrix represents the operation behavior information of a user, which is composed of multiple behavior preference features.
3. The method according to claim 1, characterized in that, The method of using the content recommendation algorithm to calculate the user's preference for the recommended content in step 2) includes: extracting the keyword features of the system article, and calculating the cosine similarity between the user's behavioral preference features and the keyword features of the system article, and using the cosine similarity as the user's preference for the recommended content.
4. The method according to claim 1, wherein The method for generating an initial recommended content queue according to the preference in step 2) includes: taking system articles whose preference exceeds a preset threshold as recommended content to generate an initial recommended content queue.
5. The method according to claim 1, characterized in that, The method of using the collaborative filtering algorithm to find similar user groups of the target user in step 3) includes: calculating the similarity between the target user and other users using the Pearson correlation coefficient based on the user's identity feature information, and selecting other users whose similarity is greater than a preset threshold as the similar user group of the target user.
6. The method according to claim 1, characterized in that The content preferred by the similar user group added to the initial recommended content queue in step 3) refers to the system articles preferred by the similar user group calculated using the content recommendation algorithm, and the article content obtained after removing the articles that the target user has read.
7. The method according to claim 1, wherein The step of calculating the target user group's preference for the recommended content in step 4) includes: Calculate the similarity sim(u, u i ) between the target user u and other users u i , where the other user u i refers to other users similar to the similar users of the target user u; Calculate the preference score score(u i , s) of other user u for the recommended content s according to the content-based recommendation algorithm i , s); Multiply sim(u, u i ) by score(u i , s), and sum over all other users u i to obtain the preference degree sim(u, s) of the target user group for the recommended content.
8. A content recommendation system based on user tag matrix, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method according to any one of claims 1 to 7 when executing the computer program.
9. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed, the steps of the method according to any one of claims 1 to 7 are implemented.