Information dissemination capability improvement method based on social relation construction

Abstract

The invention discloses an information dissemination capability improvement method based on social relation construction, which comprises the following steps: modeling user data in a social media platform into a hypergraph, extracting a high-order relation between users, evaluating user influence in combination with a PageRank algorithm and node centrality, and analyzing a result based on the influence to improve the information dissemination capability of the users. The method comprises the following steps: determining a key user having significant influence in a process of recommending specific information to a specific user, then implanting a false user, and constructing a social relationship between the false user and the key user so as to realize directional pushing of the specific information to the specific user; therefore, the visibility of the target information to the target network user is improved, and the propagation capability of the specific information to the target user is improved. According to the method, the influence of the fake user social relation on the information visibility and even the information spreading capability in the online media recommendation system is explored. Compared with other poisoning attack methods, the information spreading capability visibility improving method provided by the invention is better in performance.

Description

Technical Field

[0001] The present invention relates to the field of data security and data mining recommendation technology, and in particular to a method for improving information dissemination capabilities based on social relationship building in a social media platform. Background Art

[0002] The development of the Internet has profoundly changed the way the public obtains and shares information, making information dissemination no longer limited by traditional time and space constraints. Social platforms represented by Facebook, Twitter, WeChat and Douyin have become the main channels for information dissemination. Faced with the explosive growth of information, these social platforms use information recommendation systems based on personalized recommendation algorithms to mine users' interests and push information to users. However, due to user preferences and recommendation principles, users are deeply trapped in "information cocoons". Therefore, in most cases, specific information is simply "invisible" to specific users or user groups, which will lead to a narrow range of users' cognition in the long run, especially for growing user groups such as teenagers. How to improve the visibility of specific information to target users and ensure that specific information can reach the target audience is an important topic in current information dissemination research.

[0003] Information recommendation systems usually rely on users' historical data to generate recommendation results. However, due to the openness of recommendation systems, recommendation systems are vulnerable to data poisoning attacks. Attackers usually recruit a group of fake users to control their online information interaction behaviors, thereby introducing malicious or misleading data into the training data of the recommendation system. According to the target of the attack, poisoning attacks on recommendation systems can be divided into targeted attacks and untargeted attacks. Targeted attacks aim to increase or decrease the popularity of target items, while untargeted attacks aim to destroy the overall function of the recommendation system and deteriorate the user experience. Targeted poisoning attacks on recommendation systems can increase the visibility of specific information. Most of the existing research methods focus on improving the global visibility of a certain information through targeted poisoning attacks so that it can be pushed to as many users as possible. Moreover, these methods focus on forging information interaction records of fake users, which is suitable for recommendation systems that only rely on users' historical interaction data. For online social platforms, the social relationships between users have an important impact on the recommendation results. Generally speaking, users are more willing to receive information shared by their social circles (including friends, colleagues, and family). Therefore, these platforms usually adopt social recommendation systems. Compared with traditional recommendation systems, social recommendation systems not only consider the interests and preferences of individual users, but also the relationships and interactions between users and other users, thereby providing more accurate and personalized recommendation results. For information social recommendation systems, poisoning attacks that only forge fake user information interaction records ignore the impact of user social relationships on recommendation results.

[0004] When existing technologies forge fake user profiles through targeted poisoning attacks to increase the visibility of specific information, they do not consider the impact of social relationships between users on recommendation results. At the same time, these methods aim to increase the recommendation probability of a certain information for all users, making the attack easy to detect and lacking in targeting and concealment. How to effectively increase the visibility of specific dissemination content to specific target users, thereby improving the dissemination ability of specific information to target users, is an urgent problem to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide a method for improving information dissemination capabilities based on social relationships, so as to solve the problem that the existing technology does not fully utilize users' social relationships when improving the visibility of specific information on online social platforms to improve the dissemination capabilities of specific information to target users, and the targeting and concealment need to be improved.

[0006] In order to achieve the above tasks, the present invention adopts the following technical solutions:

[0007] A method for improving information dissemination capabilities based on social relationship construction, comprising: modeling user data in a social media platform as a hypergraph, extracting high-order relationships between users, combining the PageRank algorithm and node centrality to evaluate user influence, determining key users who have a significant impact on the process of recommending specific information to specific users based on the influence analysis results, then implanting fake users, and building social relationships between fake users and key users to achieve targeted push of specific information to specific users;

[0008] User data includes the trust relationship between users and the historical information interaction records of users.

[0009] Optionally, the social media platform includes N users and M information, where N and M are natural numbers, and there are N u trust relationships, u represents a user, and there are N trust relationships between the N users and the M information. i A historical information interaction record, i represents information, the user includes a target user a, the information includes target information b, and the social media platform recommends K information to the user through a recommendation system;

[0010] The specific methods to improve information dissemination capabilities include:

[0011] Step 1: Obtain the historical information interaction records of users and the trust relationship between users from the social media platform, and construct a hypergraph according to the preset ternary relationship model. There are 10 types of ternary relationship models, which can be divided into user-user social relationship, user-item purchase relationship and user-user-item mixed relationship; corresponding to these three types of ternary relationship models, construct 3 hypergraphs, each of which contains hyperedges composed of the ternary relationship models belonging to this type;

[0012] Step 2: Use the constructed hypergraph to train the proxy model F, where the loss function of the model F is L, and the Bayesian personalized ranking loss is usually used.

[0013]

[0014] Among them, θ is the parameter of model F, I(u) is the set of information that user u has interacted with, σ is the sigmoid function, is the model's prediction of user u's preference for information i, that is, the predicted recommendation score of user u for information i, λ is the regularization coefficient; j is the randomly sampled information that user u dislikes or is unknown;

[0015] Step 3: Use the proxy model F to predict whether the target information b appears in the K pieces of information recommended to the target user a by the model under normal circumstances, so as to determine whether it is necessary to increase the probability of recommending the target information to the target user through poisoning attacks;

[0016] Step 4: Build a fake user profile and use targeted poisoning attacks to improve the global recommendation score ranking of the target information b for the target user a among all the information, thereby increasing the probability of recommending the target information to the target user. The user profile refers to the user's social relationships and information interaction relationships.

[0017] Optionally, the step 3 specifically includes:

[0018] Step 31: Based on the historical information interaction records of the target user a and the trust relationship between users in the social network, the proxy model F is used to predict a TOP-K recommendation list L consisting of the K pieces of information that the target user a is most interested in.

[0019] Step 32: Determine whether the target information b appears in the K information recommended by the model to the target user a under normal circumstances, that is, determine whether the TOP-K recommendation list L of the target user a contains the target information b; if the target information b appears in the TOP-K recommendation list of the target user a, the visibility improvement method ends; otherwise, obtain the global recommendation score ranking of the target information b in all the information for the target user a according to the model prediction score, and execute step 4.

[0020] Optionally, the step 4 specifically includes:

[0021] Step 41, initializing a fake user according to the attack settings;

[0022] Step 42, combining the hypergraph, PageRank algorithm and node centrality to evaluate the influence of users in online social networks;

[0023] Step 43: using user influence, user trust relationship, and user historical information interaction records to select important users to build a one-way trust relationship between the fake user and the important user, thereby using the fake user's social relationship to bring the target user and target information closer together;

[0024] Step 44, after implanting the fake user, use the poisoned data to retrain the proxy model F;

[0025] Step 45, record the global recommendation score ranking of the target information b among all the information for the target user a as predicted by the proxy model after poisoning training, and determine whether the target information b appears in the K information recommended to the target user a by the model at this time, that is, determine whether the TOP-K recommendation list L of the target user a contains the target information b; if so, it means that after the attack, the target information b can enter the TOP-K recommendation list of the target user a; otherwise, it is not true.

[0026] Optionally, the step 41 includes:

[0027] Step 411: Calculate the number of fake users N to be injected according to the preset attack ratio p and the budget d of fake user social relationships. f =N·p and the number of users that the fake user needs to follow f indicates a fake user;

[0028] Step 412: Create N f fake users and initialize N f A profile of a fake user, wherein the initialization profile operation is to build a one-way trust relationship pointing to the target user and an interactive relationship pointing to the target information for the fake user.

[0029] Optionally, the step 42 includes:

[0030] Step 421: Based on the first hypergraph, that is, the hypergraph HG of user-user social relationships s The number of hyperedges corresponding to each ternary relationship pattern in , and the ternary social relationship pattern T corresponding to the type with the largest number of hyperedges is selected e , using the adjacency matrix W of the directed unweighted graph of trust relationships between online social network users and the adjacency matrix based on ternary relationships Calculating the Matrix α is used to control W and The fused hyperparameters, then the probability transfer matrix The HPR values ​​of all users are calculated using Equation 2.

[0031]

[0032] Where q is the damping coefficient, e is the unit vector, N is the total number of users of the social media platform, and HPR i is the HPR value of the i-th user, and the superscript T indicates the transpose of the matrix;

[0033] Step 422: Calculate the in-degree centrality and closeness centrality of each user in the online social network using equations 3 and 4 respectively. The in-degree centrality C in (u) and closeness center C c (u) can be used as the local structural features and global structural features of user u respectively. The in-degree centrality C of node u is combined using Equation 5. in (u) and closeness center C c (u) Get the combined centrality score C com (u);

[0034] C in (u) = in_deg(u) Formula 3;

[0035] Among them, in_deg(u) is the number of in-degree neighbors of user u, that is, the number of users who follow user u;

[0036] C c (u)=1 / ∑ v≠u d(u,v) Equation 4;

[0037] Where d(u,v) is the shortest path distance between user node u and node v, ∑ v≠u d(u,v) represents the sum of the shortest path distances from node u to all other user nodes in the online social network;

[0038]

[0039] Among them, β is the combination coefficient, maxC in and maxC c They are the maximum in-degree centrality and the maximum closeness centrality respectively;

[0040] Step 423: Combine the HPR value and the combined centrality score of each user node using Formula 6 to obtain the final user influence value INF(·). The larger the value of INF(u), the greater the influence of user u in the online social network.

[0041] INF(u)=HPR(u)+γC com(u) Formula 6;

[0042] where γ is a parameter used to control HPR and C com Hyperparameters of the ensemble.

[0043] Optionally, the step 43 includes:

[0044] Step 431: sort the users followed by the target user according to their influence values ​​INF(·) from small to large, and select s1 important users. It is a hyperparameter used to control the proportion of the number of users;

[0045] If the number of users followed by the target user is less than s1, select all users followed by the target user to get the selected user list Update the value of s1 to the total number of users followed by the target user;

[0046] Step 432: Remove The users included in the target information are sorted from small to large according to the user influence value INF(·), and s2 important users are selected. Get the selected user list If the number of users who have interacted with the target information is less than s2, then all users who have interacted with the target information are selected;

[0047] Step 433: Get the final selected user list Build a one-way trust relationship between the fake user and the selected user.

[0048] Compared with the prior art, the present invention has the following effects:

[0049] 1. Hypergraphs are used to construct social relationships between users and interactive relationships between users and information to deeply explore high-order relationships between users and information. The traditional PageRank algorithm is combined with hypergraphs and node centrality to consider the global influence of users. The visibility enhancement method provided by the present invention is more efficient and more suitable for information social recommendation systems.

[0050] 2. Comprehensively consider user influence and information popularity, and construct fake user node profiles by forging fake users' social relationships and information interaction records to increase the visibility of target information to specific users and achieve targeted dissemination of target information to specific users. At the same time, minimize the disturbance of attacks on recommendation results for non-target users and maintain the accuracy of recommendation system recommendations for non-target users. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present invention but do not constitute a limitation of the present invention. In the accompanying drawings:

[0052] Figure 1 The overall framework diagram of the method for improving information dissemination capability provided by the present invention;

[0053] Figure 2 A schematic diagram of the social relationship between users and the interactive relationship between users and information on a social media platform;

[0054] Figure 3 A ternary high-order relationship that can be constructed based on users and information in social recommendation systems;

[0055] Figure 4 To construct user-information high-level relationships;

[0056] Figure 5 This is a relationship diagram of the prediction accuracy Precision@10 of the social recommendation system after being attacked by target users with different influences under different poisoning ratios on the LastFM dataset;

[0057] Figure 6 This is a relationship diagram of the normalized discounted cumulative gain NDCG@10 of the social recommendation system after being attacked by target users with different influences at different poisoning ratios on the LastFM dataset;

[0058] Figure 7 This is a relationship diagram of the recall rate Recall@10 of the social recommendation system after being attacked by target users with different influences under different poisoning ratios on the LastFM dataset;

[0059] Figure 8 This is a relationship diagram of the change rate of the recommended ranking of target information after being attacked by target users with different influences under different poisoning ratios on the LastFM dataset;

[0060] Fig. 9 This is a graph showing the relationship between the predicted recommendation score and predicted ranking of the target user obtained by different attack methods when the poisoning ratio is 0.01 on the Yelp dataset, the target user is a high-influence user, and the target information is popular information. DETAILED DESCRIPTION

[0061] The specific contents of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments.

[0062] Social media platform: This is an online platform that allows users to create, share content and interact socially. Typical examples include Facebook, Twitter, Instagram, WeChat, TikTok, etc. Social media platforms usually provide a variety of functions, such as posting and sharing text, pictures, videos, social interactions such as comments and likes, joining groups or communities, private messaging, etc.

[0063] A social recommendation system is a part of a social media platform that is responsible for providing personalized content or social recommendations to users. A social media platform is a larger entity that includes a social recommendation system and other functions. Social recommendation system: A social recommendation system is a system that combines the user's social relationship data with traditional recommendation system technology to provide users with more personalized and accurate recommendation services. Compared with traditional recommendation systems that are mainly based on users' historical behavior data (such as browsing history, purchase history, ratings, etc.), social recommendation systems also use social relationships and social data, such as friend relationships, trust relationships, user interactions, etc. to provide more accurate and personalized recommendation results, thereby recommending content or products that are more in line with their social circles and personal interests to users, improving user satisfaction and engagement. A social recommendation system is an algorithmic system used to analyze user behavior and social relationships in order to recommend content, friends, or other services. It can generate personalized recommendations based on the user's social network, interests, historical behavior, and other data. For example, on Facebook or Twitter, you can see recommended friends or people you follow; on Instagram or TikTok, you can see recommended posts or videos.

[0064] Poisoning attacks on social recommender systems: Due to their open nature, recommender systems are vulnerable to data poisoning attacks. In data poisoning attacks on social recommender systems, attackers usually recruit a group of controlled users, control these users to strategically interact with various items, sometimes involving modifications to the functionality of certain items, and then inject the resulting malicious data into the training data of the recommender system, thereby manipulating the learning process of the recommendation model and ultimately achieving a specific goal. For example, using poisoned data for training, an e-commerce RS will provide more exposure for low-quality products. Similarly, in the context of news dissemination, false information may be deliberately recommended to certain user groups.

[0065] The method for improving information dissemination capability based on social relationship construction of the present invention aims at the problem of "incommunicability" between target information and target users caused by the "information cocoon" phenomenon. First, the data of the online information social recommendation system is modeled as a hypergraph to fully extract high-order relationships between users, and the PageRank algorithm and node centrality are combined to evaluate user influence. Based on the influence analysis results, key users with significant influence in the process of recommending specific information to specific users are determined, and then fake users are implanted, and social relationships between fake users and key users are constructed, so as to improve the visibility of target information to target network users and improve the dissemination capability of specific information to target users. The present invention explores the impact of forged user social relationships on information visibility and even information dissemination capability in online media recommendation systems. Compared with other poisoning attack methods, the method for improving the visibility of information dissemination capability proposed by the present invention has better performance.

[0066] Following the above technical solution, Figures 1 to 9 As shown, the method for improving information dissemination capabilities based on social relationship construction of the present invention utilizes a hypergraph to construct social relationships between users and interactive relationships between users and information, combines the traditional PageRank algorithm with the hypergraph and node centrality to evaluate the user's influence, and constructs a false user node profile by forging the social relationships and information interaction records of false users, thereby improving the visibility of target information to specific users and achieving targeted dissemination of target information to specific users.

[0067] Most of the existing technologies focus on attacks on recommendation systems, and there are fewer studies on attack methods for social recommendation systems. When these methods forge fake user profiles through targeted poisoning attacks to increase the visibility of specific information, they do not consider the impact of social relationships between users on recommendation results. Moreover, the target problems targeted by existing technologies usually improve the global recommendation level of a certain information or destroy the performance of the recommendation system. The corresponding real-life scenario is that driven by competition between platforms, attackers are driven by personal / commercial interests and try to manipulate recommendations to increase / decrease the ranking of specific information or damage the overall performance of the opponent's system, making the attack easy to detect and lacking in targeting and concealment.

[0068] Social recommendation systems provide personalized recommendations to users based on their social relationships and user interaction records. Generally speaking, two users with social connections are more likely to have interacted with the same information than those without social connections. Given the importance of user social relationships in social recommendation systems, this solution considers the high-order interaction relationship between users and information in social recommendation systems, selects hypergraphs to model user-user social relationships and user-information interaction relationships in the network, and uses the high-order relationships between users and information in online social networks and the influence of users to construct fake user profiles, significantly increasing the probability of recommending target information to target users.

[0069] Specific examples are given below.

[0070] like Figure 1 As shown in the figure, the social media platform includes N users and M information, N and M are natural numbers, and there are N u trust relationships, u represents a user, and there are N trust relationships between N users and M information. i interaction records, i represents information, users include target user a, information includes target information b, and K information is recommended to users through the recommendation system in the social media platform.

[0071] The method for improving information dissemination capabilities includes the following steps:

[0072] Step 1: Obtain the historical information interaction records of users and the trust relationship between users from the social media platform, and construct a hypergraph according to the preset ternary relationship model. There are 10 types of ternary relationship models, which can be divided into user-user social relationship, user-item purchase relationship, and user-user-item mixed relationship. Corresponding to these three types of ternary relationship models, three hypergraphs are constructed, each of which contains hyperedges composed of the ternary relationship models belonging to this type.

[0073] like Figure 2 As shown in the figure, on the social media platform, users interact with some information and follow some users. The solid line represents the social relationship between users, the dotted line represents the interaction relationship between users and information, and the direction of the arrow represents the trust relationship between user i and user j. When only the trust relationship between users is considered, a directed social network can be obtained.

[0074] like Figure 3As shown in Figure 1, there are two types of direct relationships between users and information in social recommendation systems, namely, social relationships between users and interactive relationships between users and information. Due to the ubiquitous existence of ternary closures in information social networks, the high-order relationship between users and information can be mined by analyzing the ternary relationship patterns between users and information. There are 10 ternary relationship patterns T here, which can be subdivided into user-user social relationships (T1-T7), user-item purchase relationships (T8), user-user-item mixed relationships (T9-T10), and user-user-item mixed relationships (T11-T11). 10 ) three categories, and thus three hypergraphs HG can be constructed s , HG c and HG h , each ternary relationship between users and between users and information corresponds to a corresponding hyperedge in the hypergraph.

[0075] Step 2: Use the constructed hypergraph to train the proxy model F, where the loss function of the model F is L, and the Bayesian personalized ranking loss is usually used.

[0076]

[0077] Among them, θ is the parameter of model F, I(u) is the set of information that user u has interacted with, σ is the sigmoid function, is the model's prediction of user u's preference for information i, that is, the predicted recommendation score of user u for information i, λ is the regularization coefficient, j is the randomly sampled information that user u dislikes or is unknown to him. Usually, the information that user u has not interacted with in the dataset is regarded as the information he dislikes or is unknown to him. The optimization goal of the loss function is to make the information i that the user likes more be ranked before the information j that the user dislikes less;

[0078] Since hypergraphs are used to model the high-order relationships between users and information in social recommendation systems, the selected proxy model should be a social recommendation model based on hypergraphs. A recommendation model with good performance can be obtained through corresponding loss functions and parameter adjustments, which can effectively simulate the results of actual social recommendation systems.

[0079] Step 3: Use the proxy model F to predict whether the target information b appears in the K information recommended to the target user a by the model under normal circumstances, so as to determine whether it is necessary to increase the probability of recommending the target information to the target user through poisoning attack.

[0080] The specific steps include:

[0081] Step 31: Based on the historical information interaction records of the target user a and the trust relationship between users in the social network, the proxy model F is used to predict a TOP-K recommendation list L consisting of the K pieces of information that the target user a is most interested in.

[0082] Step 32: Under normal circumstances, determine whether the target information b appears in the K information recommended by the model to the target user a, that is, determine whether the TOP-K recommendation list L of the target user a contains the target information b. If the target information b appears in the TOP-K recommendation list of the target user a, the visibility improvement method ends; otherwise, obtain the global recommendation score ranking of the target information b in all the information for the target user a according to the model prediction score, and execute step 4.

[0083] Step 4: Build a fake user profile and use targeted poisoning attacks to improve the global recommendation score ranking of the target information b for the target user a among all the information, thereby increasing the probability of recommending the target information to the target user. The user profile refers to the user's social relationships and information interaction relationships.

[0084] The specific steps include:

[0085] Step 41: Initialize the fake user according to the attack settings.

[0086] Optionally, the step 41 of initializing a fake user according to the attack setting includes the following steps:

[0087] Step 411: Calculate the number of fake users N to be injected according to the preset attack ratio p and the budget d of fake user social relationships. f =N·p and the number of users that the fake user needs to follow The social media platform includes N users and M information, and f represents a fake user;

[0088] Step 412: Create N f fake users and initialize N f A profile of a fake user is provided, wherein the operation of initializing the profile is to build a one-way trust relationship pointing to the target user and an interactive relationship pointing to the target information for the fake user. Note that in actual situations, it is difficult to control the attention of real users to fake users on online social platforms. Therefore, the present invention only considers building a one-way trust relationship between the fake user and the real user.

[0089] Step 42: Combine the hypergraph, PageRank algorithm and node centrality to evaluate the influence of users in the online social network; specifically, the following steps are included:

[0090] Step 421: Based on the first hypergraph, that is, the hypergraph of user-user social relationships, HG s , the number of hyperedges corresponding to each ternary relationship pattern in, and the ternary social relationship pattern T corresponding to the type with the largest number of hyperedges is selected e , using the adjacency matrix W of the directed unweighted graph of trust relationships between online social network users and the adjacency matrix based on ternary relationships Calculating the Matrix α is used to control W and The fused hyperparameters, then the probability transfer matrix The HPR values ​​of all users are calculated using Equation 2.

[0091]

[0092] Where q is the damping coefficient, e is the unit vector, N is the total number of users of the social media platform, and HPR i is the HPR value of the i-th user, where HPR is a column vector, HPR i is one of the components, and the superscript T indicates the transpose of the matrix.

[0093] Adjacency Matrix Based on Ternary Relationship middle Indicates that for a given graph, nodes i and j are in a specific ternary relationship T e The frequency of occurrence in a high-order structure (i.e., a high-order structure), the higher the frequency, the closer the relationship between nodes i and j in this relationship model.

[0094] Step 422: Calculate the in-degree centrality and closeness centrality of each user in the online social network using equations 3 and 4 respectively. The in-degree centrality represents the number of users who follow the user and is suitable for discovering local key nodes. The closeness centrality measures the inverse of the average shortest path length between a user node and other user nodes and is suitable for discovering global key nodes. in (u) and closeness center C c (u) can be used as the local structural features and global structural features of user u respectively. The in-degree centrality C of node u is combined using Equation 5. in (u) and closeness center C c (u) Get the combined centrality score C com (u).

[0095] C in (u) = in_deg(u) Formula 3;

[0096] Among them, in_deg(u) is the number of in-degree neighbors of user u, that is, the number of users who follow user u.

[0097] C c (u)=1 / ∑ v≠u d(u,v) Equation 4;

[0098] Where d(u,v) is the shortest path distance between user node u and node v, ∑ v≠u d(u,v) represents the sum of the shortest path distances from node u to all other user nodes in the online social network.

[0099]

[0100] Among them, β is the combination coefficient, maxC in and maxC c They are the maximum in-degree centrality and the maximum closeness centrality respectively.

[0101] Step 423: Use Formula 6 to combine the HPR value and the combined centrality score of each user node to obtain the final user influence value INF(·). The larger the value of INF(u), the greater the influence of user u in the online social network.

[0102] INF(u)=HPR(u)+γC com (u) Formula 6;

[0103] where γ is a parameter used to control HPR and C com Hyperparameters of the ensemble.

[0104] Step 43: Use user influence, user trust relationship, and user historical information interaction records to select important users to build a one-way trust relationship between the fake user and the important user, thereby using the fake user's social relationship to shorten the connection between the target user and the target information.

[0105] like Figure 4 As shown, after injecting fake users, fake user-user social relationships and fake user-information interactive relationships will be constructed in the system, which can specifically correspond to four of the preset ternary relationship patterns. When constructing the social relationships of fake users, specific ternary relationship patterns are generated by selecting suitable candidate users to shorten the connection between the target user and the target information, thereby affecting the recommendation results of the recommendation system for the target user.

[0106] The specific steps include:

[0107] Step 431: sort the users followed by the target user according to their influence values ​​INF(·) from small to large, and select s1 important users. is a hyperparameter used to control the proportion of the number of users; if the number of users followed by the target user is less than s1, all users followed by the target user are selected to obtain the selected user list Update the value of s1 to the total number of users followed by the target user;

[0108] Step 432: Remove The users included in the target information are sorted from small to large according to the user influence value INF(·), and s2 important users are selected. Get the selected user list If the number of users who have interacted with the target information is less than s2, then all users who have interacted with the target information are selected;

[0109] Step 433: Get the final selected user list Build a one-way trust relationship between the fake user and the selected user.

[0110] Step 44: After implanting the fake users, use the poisoned data to retrain the proxy model.

[0111] Step 45: record the global recommendation score ranking of target information b among all information for target user a predicted by the proxy model after poisoning training, and determine whether target information b appears in the K information recommended to target user a by the model at this time, that is, determine whether the TOP-K recommendation list L of target user a contains target information b. If it does, it means that after the attack, target information b can enter the TOP-K recommendation list of target user a; otherwise, it is not true.

[0112] The information visibility improvement method provided by the present invention is applied to an online social network data set to verify the performance of the advertisement push method. The selected data sets are LastFM and Yelp. The statistical results of the data sets are shown in Table 1 below.

[0113] Table 1 Experimental dataset attributes

[0114] Dataset Number of users Number of messages Number of information exchanges Number of trust relationships LastFM 1,892 17,632 92,834 25,434 Yelp 19,539 21,266 450,884 864,157

[0115] The information dissemination capability enhancement method based on social relationship construction in the social media platform provided by the present invention is compared with the random attack, popular attack and three benchmark algorithms (random selection of follow-up users, selection of follow-up users based on PageRank, and selection of follow-up users based on centrality) in the prior art. Randomattack is that the controlled users interact with the target information and some other randomly selected information, and Popular attack is that in addition to the target information, each controlled user selects 10% of the popular information and 90% of other random information. The similarities between the two are that both improve the visibility of information by constructing the information interaction records of false users, with small amount of calculation and able to improve the visibility of information to a certain extent, but ignore the social relationship between users in the social recommendation system, and it is difficult to improve the visibility of specific information to specific users in a targeted manner.

[0116] The three benchmark algorithms are: (1) RandomUser: randomly select users and build a one-way trust relationship between the fake user and the selected users; (2) PageRankUser: select important users based on PageRank and build a one-way trust relationship between the fake user and the selected important users; (3) CentralityUser: select important users based on centrality and build a one-way trust relationship between the fake user and the selected important users.

[0117] In the experiment, the number of selected messages and the number of connected users are set to be the average number of interactive messages per user and the average number of users followed by each user in the online information social network, respectively, so that the fake users are closer to the real users and the attack is more covert.

[0118] Figure 5 , Figure 6 , Figure 7 This is a performance comparison chart of the social recommendation system after attack under different poisoning ratios and different attack methods for target users with different influence on the LastFM dataset. It corresponds to three cases: low-influence users, medium-influence users, and high-influence users. The corresponding indicators are prediction accuracy Precision@K, normalized discounted cumulative gain NDCG@10, and recall Recall@K, which reflect the recommendation performance of the social recommendation system. The larger the indicator value, the better the recommendation effect of the recommendation system. The target information is information with low popularity, and K is 10. Figure 8 The relative change rate between the ranking of target information in the target user's global recommendation list and the original ranking after the poisoning attack is shown for target users with different influences on the LastFM dataset under different poisoning ratios and different attack methods. Fig. 9 The effect of increasing the visibility of popular information to high-influence users through poisoning attacks in the Yelp dataset is given.

[0119] like Figure 5 , Figure 6 , Figure 7As shown, under target users with different influences and poisoning ratios, the method of the present invention has shown relatively excellent results. In the scenario of low-influence target users, the method of the present invention reaches the peak effect when the poisoning ratio is 0.01, and then slightly decreases but still maintains a high level. For target users with medium influence, the method of the present invention shows high NDCG@10 values ​​at multiple poisoning ratios, especially when the poisoning ratio is 0.03. In the scenario of high-influence target users, the method of the present invention is also better than other methods in overall effect, especially when the poisoning ratio is low. The above results show that the method of the present invention can better maintain the normal performance of the social recommendation system and will not cause serious damage to the overall performance of the recommendation system.

[0120] The reason is that the Random attack and Popular attack methods focus on the relationship between users and information, and only consider information interaction behaviors, but do not consider the social relationship between users in the network; while the three benchmark methods focus on building social relationships between fake users and real users, and fail to fully consider the social influence of users and the high-order relationship between users and information. This further proves that the present invention is a method that fully considers the social influence of users and the high-order relationship between users and information.

[0121] like Figure 8 As shown, under the setting of target users with different influences, the method of the present invention shows significant advantages when the poisoning ratio is small. As the poisoning ratio increases, the improvement effect tends to be stable, but the overall effect is still better than other benchmark methods. In the scenario of low-influence target users, the method of the present invention maintains a deltarank@10 value close to 100% at each poisoning ratio, indicating that the target information recommendation ranking improvement effect is significant, and the visibility of the target information to the target user is significantly increased, and even enters the TOP-10 recommendation list. For target users with medium influence, the method of the present invention outperforms other benchmark methods at most poisoning ratios, especially when the poisoning ratio is 0.03. In the scenario of high-influence target users, the method of the present invention also shows a high ranking stability, especially when the poisoning ratio is high, its deltarank@10 value remains at a high level.

[0122] The reason is that the method of the present invention comprehensively considers the local influence and global influence of users and the high-order relationship between users and information, and constructs a ternary high-order relationship related to the target information to bring the target information and the target user closer together. When the value of deltarank@10 is close to 100%, it means that the visibility of the target information to the target user is significantly improved and has even entered the target user's TOP-10 information recommendation list. Therefore, as the poisoning ratio increases, the push effect of the method of the present invention tends to stabilize instead of continuing to improve significantly.

[0123] like Fig. 9 As shown, for the Yelp dataset, when the poisoning ratio is 0.01, the target user is a high-influence user, and the target information is popular information, compared with the three benchmark methods, the method of the present invention obtains the highest predicted recommendation score for the target information for the target user, and ranks best in the global recommendation list of the target user.

[0124] The reason is that, when selecting users to follow for a fake user, the method of the present invention not only considers the local influence and global influence of the user, but also considers the high-order relationship between the user and the information. By selecting key users to construct a ternary high-order relationship related to the target information, the association between the target information and the target user is shortened.

Claims

1. A method for improving information dissemination capabilities based on social relationship building, characterized in that: include: The user data in the social media platform is modeled as a hypergraph, and the high-order relationships between users are extracted. The user influence is evaluated by combining the PageRank algorithm and node centrality. Based on the influence analysis results, the key users who have a significant impact on the process of recommending specific information to specific users are determined. Then, fake users are implanted, and the social relationship between fake users and key users is constructed to achieve targeted push of specific information to specific users. User data includes the trust relationship between users and the historical information interaction records of users.

2. The method for improving information dissemination capability based on social relationship construction according to claim 1, characterized in that: The social media platform includes N users and M information, where N and M are natural numbers, and there are N u trust relationships, u represents a user, and there are N trust relationships between the N users and the M information. i A historical information interaction record, i represents information, the user includes a target user a, the information includes target information b, and the social media platform recommends K information to the user through a recommendation system; The specific methods to improve information dissemination capabilities include: Step 1: Obtain the historical information interaction records of users and the trust relationship between users from the social media platform, and construct a hypergraph according to the preset ternary relationship model. There are 10 types of ternary relationship models, which can be divided into user-user social relationship, user-item purchase relationship and user-user-item mixed relationship; corresponding to these three types of ternary relationship models, construct 3 hypergraphs, each of which contains hyperedges composed of the ternary relationship models belonging to this type; Step 2: Use the constructed hypergraph to train the proxy model F, where the loss function of the model F is L, and the Bayesian personalized ranking loss is usually used. Among them, θ is the parameter of model F, I(u) is the set of information that user u has interacted with, σ is the sigmoid function, is the model's prediction of user u's preference for information i, that is, the predicted recommendation score of user u for information i, λ is the regularization coefficient; j is the randomly sampled information that user u dislikes or is unknown; Step 3: Use the proxy model F to predict whether the target information b appears in the K pieces of information recommended to the target user a by the model under normal circumstances, so as to determine whether it is necessary to increase the probability of recommending the target information to the target user through poisoning attacks; Step 4: Build a fake user profile and use targeted poisoning attacks to improve the global recommendation score ranking of the target information b for the target user a among all the information, thereby increasing the probability of recommending the target information to the target user. The user profile refers to the user's social relationships and information interaction relationships.

3. The method for improving information dissemination capability based on social relationship construction as claimed in claim 2, characterized in that: The step 3 specifically includes: Step 31: Based on the historical information interaction records of the target user a and the trust relationship between users in the social network, the proxy model F is used to predict a TOP-K recommendation list L consisting of the K pieces of information that the target user a is most interested in. Step 32: Determine whether the target information b appears in the K information recommended by the model to the target user a under normal circumstances, that is, determine whether the TOP-K recommendation list L of the target user a contains the target information b; if the target information b appears in the TOP-K recommendation list of the target user a, the visibility improvement method ends; otherwise, obtain the global recommendation score ranking of the target information b in all the information for the target user a according to the model prediction score, and execute step 4.

4. The method for improving information dissemination capability based on social relationship construction according to claim 2 or 3, characterized in that: The step 4 specifically includes: Step 41, initializing a fake user according to the attack settings; Step 42, combining the hypergraph, PageRank algorithm and node centrality to evaluate the influence of users in online social networks; Step 43: using user influence, user trust relationship, and user historical information interaction records to select important users to build a one-way trust relationship between the fake user and the important user, thereby using the fake user's social relationship to bring the target user and target information closer together; Step 44, after implanting the fake user, use the poisoned data to retrain the proxy model F; Step 45, record the global recommendation score ranking of the target information b among all the information for the target user a predicted by the proxy model after poisoning training, and determine whether the target information b appears in the K information recommended to the target user a by the model at this time, that is, determine whether the TOP-K recommendation list L of the target user a contains the target information b; if it does, it means that after the attack, the target information b can enter the TOP-K recommendation list of the target user a; otherwise, it is not true.

5. The method for improving information dissemination capability based on social relationship construction as claimed in claim 4, characterized in that: The step 41 comprises: Step 411: Calculate the number of fake users N to be injected according to the preset attack ratio p and the budget d of fake user social relationships. f =N·p and the number of users that the fake user needs to follow f indicates a fake user; Step 412: Create N f fake users and initialize N f A profile of a fake user, wherein the initialization profile operation is to build a one-way trust relationship pointing to the target user and an interactive relationship pointing to the target information for the fake user.

6. The method for improving information dissemination capability based on social relationship construction according to claim 4, characterized in that: The step 42 comprises: Step 421: Based on the first hypergraph, that is, the hypergraph HG of user-user social relationships s The number of hyperedges corresponding to each ternary relationship pattern in , and the ternary social relationship pattern T corresponding to the type with the largest number of hyperedges is selected e , using the adjacency matrix W of the directed unweighted graph of trust relationships between online social network users and the adjacency matrix based on ternary relationships Calculating the Matrix α is used to control W and The fused hyperparameters, then the probability transfer matrix The HPR values ​​of all users are calculated using Equation 2. Where q is the damping coefficient, e is the unit vector, N is the total number of users of the social media platform, and HPR i is the HPR value of the i-th user, and the superscript T indicates the transpose of the matrix; Step 422: Calculate the in-degree centrality and closeness centrality of each user in the online social network using equations 3 and 4 respectively. The in-degree centrality C in (u) and closeness center C c (u) can be used as the local structural features and global structural features of user u respectively. The in-degree centrality C of node u is combined using Equation 5. in (u) and closeness center C c (u) Get the combined centrality score C com (u); C in (u) = in_deg(u), Equation 3; Among them, in_deg(u) is the number of in-degree neighbors of user u, that is, the number of users who follow user u; C c (u) = 1 / ∑ v≠u d(u, v) Equation 4; Where d(u,v) is the shortest path distance between user node u and node v, ∑ v≠u d(u,v) represents the sum of the shortest path distances from node u to all other user nodes in the online social network; Among them, β is the combination coefficient, maxC in and maxC c They are the maximum in-degree centrality and the maximum closeness centrality respectively; Step 423: Combine the HPR value and the combined centrality score of each user node using Formula 6 to obtain the final user influence value INF(·). The larger the value of INF(u), the greater the influence of user u in the online social network. INF(u) = HPR(u) + γC com (u) Equation 6; where γ is a parameter used to control HPR and C com Hyperparameters of the ensemble.

7. The method for improving information dissemination capability based on social relationship construction according to claim 4, characterized in that: The step 43 comprises: Step 431: sort the users followed by the target user according to their influence values ​​INF(·) from small to large, and select s1 important users. It is a hyperparameter used to control the proportion of the number of users; If the number of users followed by the target user is less than s1, select all users followed by the target user to get the selected user list Update the value of s1 to the total number of users followed by the target user; Step 432: Remove The users included in the target information are sorted from small to large according to the user influence value INF(·), and s2 important users are selected. Get the selected user list If the number of users who have interacted with the target information is less than s2, then all users who have interacted with the target information are selected; Step 433: Get the final selected user list Build a one-way trust relationship between the fake user and the selected user.

Images