A relationship prediction method based on dynamic interaction of social network platforms
By constructing a hierarchical progressive user interaction matrix sequence and text feature representation in social networks, combining the LSTM network and graph attention model, the problem of failing to fully consider the hidden information and progressive order of interaction behavior in the prior art is solved, and more accurate social network link prediction and user relationship analysis are achieved.
Patent Information
- Application Number
- CN202210972715.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-15
- Publication Date
- 2025-07-01
- Estimated Expiration
- 2042-08-15
AI Technical Summary
The prior art fails to fully consider the hidden information and progressive order between interactive behaviors when linking predictions in social networks, resulting in poor prediction results.
A relationship prediction method based on dynamic interaction of social network platforms is proposed. By constructing a hierarchical progressive user interaction matrix sequence and text feature representation, combining LSTM network and graph attention model, integrating user characteristics and interactive network structure characteristics, calculating the correlation and similarity between user nodes, and performing relationship prediction.
This method can more accurately predict social network links and analyze the types of links established between users, improving the prediction effect and understanding of social rules.
Smart Images

Figure CN115293437B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of social network relationship prediction, and specifically relates to a relationship prediction method based on dynamic interaction of a social network platform. Background Art
[0002] Relationship prediction in social networks refers to using existing links to predict the possibility of future interactions and establishing links between users. Establishing a link actually represents the interaction behavior generated between users in the social network neighborhood, that is, social behavior. Traditionally, interactions between people usually occur at a certain determined time and space. Only when two people meet can there be social behavior. The social circle is small and social behavior is scarce. Nowadays, with the rise of social networks such as Twitter, Weibo, and WeChat, people have abandoned the distance limit of traditional social interactions and can become friends with people thousands of miles away through online social platforms, which is of great significance to all aspects of people's lives. With the development speed of online social platforms, in order to help social platforms retain users and make more accurate friend recommendations, link prediction has become an important research direction in social networks. In the research of traditional link prediction, the impact of hidden information existing between interaction behaviors on the evolution of the entire social network is not considered. Therefore, carrying out research on different interaction behaviors and the hidden information between interaction behaviors plays an important role in social network information mining and prediction.
[0003] In recent years, many scholars have conducted a large number of studies on link prediction models, mainly based on dynamic network representation learning methods such as DynGEM and DDNE, and static network representation learning methods such as node2vec and GCN. The dynamic graph link prediction model based on DynGEM has the following advantages: as time goes by, the embedding is stable; it can handle increasing dynamic graphs; it is more efficient than directly applying static graph methods to each snapshot. DDNE uses GRU as an encoder to capture the time information in the dynamic network, so as to perform link prediction in the dynamic network. Based on the node2vec static network representation learning model, it uses random walks to capture the proximity in the network and maps all nodes to a low-dimensional representation space that maintains proximity. GCN constructs a semi-supervised node embedding model, and by encoding the network topology structure and network node features, it obtains node representations containing rich information. Although many scholars have done well in the field of dynamic network link prediction, in terms of interaction rules, no one has captured this potential feature by constructing a hierarchical progressive interaction network. Even when constructing graph snapshots, only the factor of time is considered, and the progressive order between interaction behaviors is not considered. This article conducts in-depth research on the subtle social rules of humans. Summary of the Invention
[0004] To solve the above problems, the present invention provides a relationship prediction method based on dynamic interaction of a social network platform, including the following steps:
[0005] S1. Online obtain user data, which includes interaction information and text information of users in the social network;
[0006] S2. Process the text information using a negative sampling model to obtain a word sequence, construct a text sequence through the word sequence, and input the text sequence into an LSTM network to obtain a text feature representation;
[0007] S3. Construct a hierarchical and progressive user interaction matrix sequence through the interaction information, perform convolution on it with the text feature representation to obtain a hierarchical and progressive user-text interaction snapshot graph sequence, and input the user-text interaction snapshot graph sequence into an LSTM network to obtain an interaction network structure feature representation;
[0008] S4. Fuse the text feature representation and the interaction network structure feature representation to obtain a user feature representation;
[0009] S5. Calculate the correlation between the user node u i in the user feature representation and the remaining user nodes, and perform a descending order arrangement. Select the top K user nodes corresponding to the correlation from the sequence to form a strong correlation user group of the user node u i ;
[0010] S6. Input the strong correlation user group of the user node u i into a graph attention model to obtain new features of the user node u i ;
[0011] S7. Repeat steps S5 - S6 to obtain new features of each user node in the social network, calculate the similarity between any user node and the remaining user nodes based on the new features, and obtain the relationship prediction result of each user node.
[0012] Further, the specific process of obtaining the text feature representation in step S2 is as follows:
[0013] S210. Obtain the text information of N users, that is, all original content and forwarded content posted by each user on the social network in the most recent month;
[0014] S211. Process the text information through a tokenizer to construct a user dictionary for each user, and encode each word in the user dictionary; input the encoded word into a negative sampling model to obtain the word vector of the word;
[0015] S212. The word vectors corresponding to all words in a piece of content form a word sequence; divide it into an original word sequence and a forwarded word sequence according to the categories of the original content and the forwarded content;
[0016] S213. Select s original word sequences or s reposted word sequences of user i, and form a time-series text sequence of user i in chronological order The time-series text sequences of N users form a text vector matrix C t ;
[0017] S214. Pass the text vector matrix C t through the LSTM network to generate a text feature representation C t+1 ;
[0018] Among them, is the text vector matrix composed of the time-series text sequences of N users, represents the time-series text sequence of the i-th user at time t, represents the s-th word sequence in the time-series text sequence of the i-th user at time t.
[0019] Furthermore, the process of constructing a hierarchical and progressive user interaction matrix sequence in step S3 through mutual information is as follows:
[0020] S310. Obtain data on various interaction behaviors of users. For each interaction behavior, construct an adjacency matrix, and arrange the adjacency matrices in ascending order of the difficulty of generating relationship links according to the interaction behaviors to obtain an adjacency matrix sequence;
[0021] S311. Perform a forward summation on the adjacency matrix sequence to obtain a hierarchical and progressive user interaction matrix sequence, denoted as:
[0022]
[0023] Among them, A i represents the i-th adjacency matrix in the adjacency matrix sequence, n represents the number of layers of the hierarchical and progressive user interaction matrix sequence, and M n represents the progressive interaction matrix of the n-th layer in the hierarchical and progressive user interaction matrix sequence.
[0024] Furthermore, step S5 uses cosine similarity to calculate the correlation between user nodes, denoted as:
[0025]
[0026] Among them, u x and u y are any two user nodes in the user feature representation, V(u x ) and V(u y ) respectively represent the initial feature vectors of user node u x and user node u y , and p i , q iDenote the components of the preliminary feature vectors V(u x ) and V(u y ), and b represents the dimension of the preliminary feature vector.
[0027] Furthermore, the relationship prediction model outputs the prediction result through the probability (similarity) of the user node u i establishing a connection with its neighbor nodes, which is expressed as:
[0028]
[0029] where, represents the correlation between the user nodes u i and u j , represents the strongly correlated user group of the user node u i , Y = {0, 1, 2, 3, 4} represents that the predicted link strength levels are 0, 1, 2, 3, 4 respectively, and the strength levels from small to large represent the link relationships established between users as no link, like, comment, forward, and follow. U, L, C, Fd, and Fw respectively represent the value ranges of the link strength levels 0, 1, 2, 3, and 4.
[0030] Furthermore, during the training process of the relationship prediction model, the loss function is used to calculate the loss, which is expressed as:
[0031]
[0032] where, P y represents the similarity probability of the user under different connection types, and Y represents the predicted link strength level.
[0033] Advantages of the present invention:
[0034] The present invention proposes a relationship prediction method based on dynamic interaction in social networks, introducing the potential relationships between interaction behaviors. It can not only perform more accurate social network link prediction but also analyze the types of links established between users. It mainly extracts the communication rules followed by humans in the interaction network by constructing a hierarchical progressive interaction matrix. Most previous scholars extracted the dynamic features of social networks by constructing a time-series link prediction model. The present invention records all changes on a single graph and then divides the graph through various interaction behaviors to extract the network structure diagrams of different behaviors. Specifically, in the embodiment, Weibo data is used for experimental testing. For 28 days within a month, all data of the first 27 days is recorded on a single graph, and then the graph is divided through various interaction behaviors to extract the network structure diagrams of different behaviors for subsequent prediction. Considering the progressive relationship of interaction behaviors, the adjacency matrix is processed by forward summation, enabling more important interaction features to be extracted after passing through the LSTM network. To enhance the representation of user nodes in the present invention, the present invention reconstructs the new features of nodes using strongly correlated user groups and attention networks to improve the prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a flowchart of an embodiment of the present invention;
[0036] Figure 2 is a schematic diagram showing the hidden information of the TT2vec algorithm in an embodiment of the present invention;
[0037] Figure 3 is a schematic diagram of HPIN2vec in an embodiment of the present invention;
[0038] Figure 4 is a link prediction model diagram of an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0040] In one embodiment, as Figure 1 、 Figure 4 shown, a relationship prediction method based on dynamic interaction on a social network platform includes the following steps:
[0041] S1. Online acquisition of user data:
[0042] The ways to obtain user data are not limited to one. It can be obtained from public data websites or by using mature social network public APIs. What needs to be obtained here is the text data (text information) of user forwards and originals, as well as the network structure data formed under different interaction behaviors, that is, historical behavior data (interaction information). The text data needs to obtain the time of text publication or forwarding for subsequent temporal text processing of TT2vec.
[0043] Specifically, after obtaining the original user data, data cleaning is performed on it. Usually, the original data obtained is unstructured and cannot be directly used for data analysis. Through simple data cleaning, most unstructured data can be structured, so that outliers or null values no longer appear, reducing the inconvenience brought to subsequent calculations. The cleaned user data is stored in a database, and the data is further standardized through the table structure, while improving the data retrieval efficiency and the mapping of inter-table relationships.
[0044] S2. Obtain user interest discovery and social rule discovery:
[0045] For user interest discovery, in this embodiment, interest factors that affect users to establish links are dynamically extracted from the text information published and forwarded by users in the social network. By designing the TT2vec (Timing Text to vector) algorithm, the hidden dynamic information between users and the published text is discovered, the text information network is calculated, and the current and subsequent interest preference information of users is analyzed;
[0046] For social rule discovery, in this embodiment, a hierarchical progressive user interaction matrix is introduced, and a GCN model based on text features is constructed. In order to discover the potential rules of user interactions, the evolution mode of the interaction network is further captured through TT2vec.
[0047] In one embodiment, the TT2vec algorithm is designed to construct the hidden information between users and their affiliated texts, thereby calculating the text information network and realizing user interest discovery. As Figure 2 shown, it includes: processing the text information by using a negative sampling model to obtain a word sequence, constructing a temporal text sequence through the word sequence, and inputting the temporal text sequence into an LSTM network to obtain a text feature representation (text information network).
[0048] Specifically, taking the Weibo social network platform as an example, the complete steps are:
[0049] S210. Obtain the text information of N users, that is, the texts of all original Weibo posts and forwarded Weibo posts published by each user on the Weibo social network in the most recent month;
[0050] S211. Process the text information through a tokenizer to construct a user dictionary for each user, and encode each word in the user dictionary through index-word, word-index, and one-hot; input the encoded words into the negative sampling model to obtain the word vectors of these words, and the dimension of the word vectors is l;
[0051] Specifically, the negative sampling model adopted in this embodiment is a skip-gram model based on negative sampling. In order to reduce the influence of other noise words unrelated to the input word (the encoded word in this embodiment) on the output (the word vector of each word in the user dictionary in this embodiment) and the computational complexity, its objective function is expressed as:
[0052]
[0053] where M and neg respectively represent the total number of words in the user dictionary and the total number of negative sampling words, and σ(·) represents the sigmoid function, represents the input vector (encoded word w) of the i-th word w in the user dictionary, represents the context output vector and the noise word output vector.
[0054] S212. The word vectors corresponding to all words in the text of a Weibo form a word sequence; divide them into an original word sequence and a reposted word sequence according to the categories of the original Weibo and the reposted Weibo;
[0055] S213. Considering the redundancy and timeliness of text information in the social network, select s original word sequences or s reposted word sequences of user i, and form a time-series text sequence of user i in chronological order The time-series text sequences of N users form a text vector matrix
[0056] Specifically,
[0057] where, represents the time-series text sequence of the i-th user at time t, represents the s-th word sequence in the time-series text sequence of the i-th user at time t, and m (padded with zeros if insufficient) is the length of the text sorted by time for the user.
[0058] S214. Pass the text vector matrix C t through the LSTM network to generate a text feature representation C t+1 ;
[0059] Since the user's interest preferences are dynamic, over time, randomly formed interests will be forgotten, long-term formed interests will be retained, and currently formed interests indicate the current author's interest tendency. Therefore, by using the long-term dependence function of the LSTM model, the user's long-term interest characteristics and current interest characteristics are extracted, and the random interests in the process are attenuated, so as to improve the ability to express the user's preferences, obtain a dynamic representation of the user's interest preferences at the next moment, and better serve link prediction:
[0060]
[0061] Among them, represents the representation at time t+1 generated after the time series text sequence passes through the LSTM filter,
[0062] In one embodiment, the hierarchical progressive user interaction network representation algorithm HPIN2vec (Hierarchical Progressive Interactive Network to vector) is introduced to mine the potential regular features between interaction behaviors, calculate user interaction behaviors, and realize social rule discovery. For example, Figure 3 as shown, it includes: constructing a hierarchical progressive user interaction matrix sequence through interaction information, performing convolution with the text feature representation to obtain a hierarchical progressive user text interaction snapshot graph sequence, and inputting the user text interaction snapshot graph sequence into the LSTM network to obtain the interaction network structure feature representation.
[0063] Specifically, taking the Weibo social network platform as an example, the complete steps are as follows:
[0064] S220. Obtain data of various user interaction behaviors, including data of four interaction behaviors: like, comment, forward, and follow in this embodiment;
[0065] S221. Construct an adjacency matrix for each interaction behavior, and arrange the adjacency matrices in ascending order of the difficulty of generating relationship links according to the interaction behavior to obtain an adjacency matrix sequence; in this embodiment, the difficulty of generating relationship links for the four interaction behaviors from small to large is like, comment, forward, and follow;
[0066] S222. Since different interaction behaviors have different difficulties in generating relationship links, there is a shortcoming that the number of links learned by a simple behavior progression sequence is getting fewer and fewer. Therefore, the adjacency matrix sequence is summed up front to obtain a hierarchical progressive user interaction matrix sequence M t , and the calculation of each layer's progressive interaction matrix is expressed as:
[0067]
[0068] Among them, A i represents the i-th adjacency matrix in the sequence of adjacency matrices, n represents the number of layers of the hierarchical progressive user interaction matrix sequence, and n is the same as the number of types of interaction behaviors. In this embodiment, there are four interaction behaviors, so n = 4, M n represents the progressive interaction matrix of the n-th layer in the hierarchical progressive user interaction matrix sequence, M t ={M1, M2, …, M n}.
[0069] Specifically, A1, A2, A3, and A4 respectively represent the adjacency matrices of like, comment, forward, and follow relationships. Then M1 = A1, M2 = A1 + A2, M3 = A1 + A2 + A3, and M4 = A1 + A2 + A3 + A4.
[0070] S223. In order to further process the interaction features other than the user's own information and construct a unified expression matrix, and in order to only obtain the features of interaction behaviors without attenuating the features of the user's own information, the degree matrix in the graph convolutional network is removed, and a graph convolutional network model based on TT2vec is constructed; the text feature representations of all users are used as the input of the graph convolutional network model, and they are respectively convolved with the progressive interaction matrices in the hierarchical progressive user interaction matrix sequence M t ={M1, M2, …, M n} to obtain a sequence of hierarchical progressive user text interaction snapshot graphs, which can be simply expressed as:
[0071]
[0072] G1 represents the first convolution result in the sequence of user text interaction snapshot graphs, which is obtained by convolving M1 with C t+1 .
[0073] S224. In order to obtain the evolution information of the network structure, the sequence of user text interaction snapshot graphs is input into the LSTM network for information extraction to obtain the representation of the interaction network structure features, which is expressed as:
[0074]
[0075] S3. The text feature representation and the interaction network structure feature representation are fused to obtain the user feature representation:
[0076]
[0077] V t+1 ={V(u0), ···, V(u i ), …, V(u N ), u i represents the i-th user node, V(ui ) represents the user node u i 's initial feature vector.
[0078] S4. Calculate the correlation between the user node u i in the user feature representation and the remaining user nodes, and sort them in descending order. Select the top k user nodes corresponding to the correlation from the sequence to form the strongly correlated user group of the user node u i
[0079] The correlation between users is obtained by calculating the similarity between node vectors. The higher the similarity between nodes, the stronger the correlation.
[0080] Specifically, the cosine similarity is used to calculate the correlation between user nodes, which is expressed as:
[0081]
[0082] where, u x , u y are any two user nodes in the user feature representation, V(u x ), V(u y ) respectively represent the initial feature vectors of the user node u x and the user node u y , p i , q i represent the components of the initial feature vectors V(u x ) and V(u y ), and b represents the dimension of the initial feature vector.
[0083] S5. Input the strongly correlated user group of the user node u i into the graph attention model GAT to obtain the new feature V(u i )′ of the user node u i .
[0084] Specifically,
[0085]
[0086]
[0087] To enable self-attention to stably represent nodes, a multi-head attention mechanism is introduced here to improve the representation ability of the model:
[0088]
[0089] where, θ ij represents the user u i and the user u jThe normalized attention coefficient, Att ij represents user u i and user u j the attention coefficient between them, represents a weight vector of size 2l′, W represents the l×l′ transformation matrix, represents the user node u after obtaining the strongly correlated user group i the eigenvector of, represents the user node u i user u in the strongly correlated user group of j the eigenvector of, R represents the R-fold attention mechanism.
[0090] S7. Repeat steps S5 - S6 to obtain the new features of each user node in the social network, calculate the similarity between any user node and the remaining user nodes based on the new features, and obtain the relationship prediction result of each user node.
[0091] Specifically, as Figure 4 shown, this embodiment defines the relationship link prediction task as a multi-classification task, and transforms the interactive behavior of establishing connections from individuals to groups, and then analyzes the social activity link trend to determine the link type. During the prediction process, the activation function softmax is used to represent the output as the probability value of whether a link is established between users. The 5 nodes output by the neural network respectively represent no connection, like connection, comment connection, forward connection, and follow connection. The activation function of each node is defined as follows:
[0092]
[0093] Among them, represents the probability of the i-th type of relationship link, n represents the number of types of link types, and σ i (z) represents the normalized probability value.
[0094] The final output value of the relationship prediction model represents the link strength level of establishing a connection between user u i and its neighbors. Whether a link is established and the type of established link are distinguished according to the link strength level. The specific definition is as follows:
[0095]
[0096] Among them, represents the correlation between user node u i and u j between them, represents the strongly correlated user group of user node u i , Y = 0 represents that the predicted link strength level is 0, and user node u i and u jThere may be no link relationship; Y = 1 indicates that the predicted link strength level is 1, and for user node u i and u j there may be a like link relationship established; Y = 2 indicates that the predicted link strength level is 2, and for user node u i and u j there may be a comment link relationship established; Y = 3 indicates that the predicted link strength level is 3, and for user node u i and u j there may be a forward link relationship established; Y = 4 indicates that the predicted link strength level is 4, and for user node u i and u j there may be a follow link relationship established. U, L, C, Fd, Fw are hyperparameters, representing the value ranges for link strength levels 0, 1, 2, 3, and 4 respectively.
[0097] Specifically, during the training process of the relationship prediction model, a loss function is used to calculate the loss, which is expressed as:
[0098] L = -∑ y∈Y P y log(P y )
[0099] where P y represents the similarity probability of the user under different connection types, and Y represents the predicted link strength level.
[0100] In the present invention, unless otherwise clearly defined and limited, terms such as "installation", "setting", "connection", "fixation", "rotation", etc. shall be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components or the interaction relationship between two components. Unless otherwise clearly defined, for those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0101] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A relationship prediction method based on dynamic interaction of a social network platform, characterized in that It includes the following steps: S1. Obtain user data online, which includes the interaction information and text information of the user in the social network; S2. Process the text information using a negative sampling model to obtain a word sequence, construct a text sequence through the word sequence, and input the text sequence into an LSTM network to obtain a text feature representation; S3. Construct a hierarchical and progressive user interaction matrix sequence through the interaction information, perform convolution on it and the text feature representation to obtain a hierarchical and progressive user-text interaction snapshot graph sequence, and input the user-text interaction snapshot graph sequence into an LSTM network to obtain an interaction network structure feature representation; The process of constructing a hierarchical and progressive user interaction matrix sequence through the interaction information in step S3 is as follows: S310. Obtain data on various interaction behaviors of the user, construct an adjacency matrix for each interaction behavior, arrange the adjacency matrices in ascending order according to the difficulty of generating relationship links by the interaction behavior to obtain an adjacency matrix sequence; S311. Perform forward summation on the adjacency matrix sequence to obtain a hierarchical and progressive user interaction matrix sequence, denoted as: Among them, A i represents the i-th adjacency matrix in the sequence of adjacency matrices, n represents the number of layers of the hierarchical progressive user interaction matrix sequence, and M n represents the progressive interaction matrix of the n-th layer in the sequence of hierarchical progressive user interaction matrices; S4. Fuse the text feature representation and the interaction network structure feature representation to obtain a user feature representation; S5. Calculate the correlation between the user node u in the user feature representation and the remaining user nodes, and sort them in descending order. Select the top K user nodes corresponding to the correlation from the sequence to form the strongly correlated user group of the user node u i ; i Step S5 uses cosine similarity to calculate the correlation between user nodes, denoted as: Among them, u x and u y are any two user nodes in the user feature representation. V(u x ) and V(u y ) respectively represent the preliminary feature vectors of user node u x and user node u y . p i and q i represent the components of the preliminary feature vectors V(u x ) and V(u y ). b represents the dimension of the preliminary feature vector; S6. Input the strongly relevant user group of user node u i into the graph attention model to obtain the new features of user node u i ; S7. Repeat steps S5 - S6 to obtain new features of each user node in the social network, calculate the similarity between any user node and the remaining user nodes based on the new features, and obtain the relationship prediction result of each user node.
2. The relationship prediction method based on dynamic interaction of a social network platform according to claim 1, characterized in that The specific process of obtaining the text feature representation in step S2 is as follows: S210. Obtain the text information of N users, that is, all original content and forwarded content published by each user on the social network in the most recent month; S211. Process the text information through a tokenizer to construct a user dictionary for each user, and encode each word in the user dictionary; input the encoded word into the negative sampling model to obtain the word vector of the word; S212. The word vectors corresponding to all words in a piece of content form a word sequence; it is divided into an original word sequence and a forwarded word sequence according to the categories of the original content and the forwarded content; S213. Select s original word sequences or s retweet word sequences of user i, and form a time-series text sequence of user i in chronological order The time-series text sequences of N users form a text vector matrix C t ; S214. Generate the text feature representation C t from the text vector matrix C t+1 ; Among them, is a text vector matrix composed of time-series text sequences of N users, represents the time-series text sequence of the i-th user at the t-th moment, represents the s-th word sequence in the time-series text sequence of the i-th user at the t-th moment.
3. A relationship prediction method based on dynamic interaction of a social network platform according to claim 1, characterized in that The relationship prediction model outputs the prediction result through the probability that the user node u i establishes a connection with its neighbor nodes, expressed as: Among them, represents the user node u i and u j the correlation between them, represents the strongly correlated user group of the user node u i Y = {0, 1, 2, 3, 4} represents that the predicted link strength levels are 0, 1, 2, 3, and 4 respectively, and U, L, C, Fd, and Fw respectively represent the value ranges of the link strength levels of 0, 1, 2, 3, and 4.
4. A relationship prediction method based on dynamic interaction of a social network platform according to claim 1, characterized in that In the training process of relationship prediction, a loss function is used to calculate the loss, which is denoted as: L = -∑ y∈Y P y log(P y ) Among them, P y represents the similarity probability of the user under different connection types, and Y represents the predicted link strength level.