Social network user recommendation method and system

Abstract

The invention discloses a social network user recommendation method and system. Relates to the field of social network systems. According to the social network user recommendation method and system, the social network user recommendation method and system for analysis and prediction based on user vectors, multiple behavior vectors and interests are provided specially for scenes of a social network platform, and according to collected behavior information data of the user vectors, a deep neural network model is adopted for analysis and learning; the method comprises the following steps: establishing a cascade relationship of various behavior vectors of user vectors by using a graph convolutional network, modeling influence values of different factor vectors on different user preference scores through a decoupling representation technology, screening potential friends through two stages of Pearson similarity calculation and a hierarchical clustering method, and recommending the friends. The method aims to accurately recognize and recommend potential social relations which a user vector may be interested in through data modeling, so that the user vector experience and the activeness of a social network platform are improved.

Description

Technical Field

[0001] The present invention relates to a social network system, and in particular to a social network user recommendation method and system. Background Art

[0002] With the popularity of social networks, users are faced with more and more information and contact options, but how to effectively discover and establish social relationships related to their interests has become a challenge. Traditional social networking platforms rely on users to add friends or follow accounts on their own, and mostly make recommendations based on the interests selected by users. There is a lack of recommendation mechanisms for user behavior analysis, which limits the quality of user experience and social interaction.

[0003] User recommendations can cover a variety of scenarios and situations. The main purpose is to quickly recommend relevant accounts to users in areas such as hobbies, learning resources, or life and work, so as to broaden the user's social relationships. Therefore, it is very important to ensure the reliability of user recommendations. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a method and system for recommending users of a social network. The technical problem to be solved is that traditional social network platforms rely on users to add friends or follow accounts on their own and lack an accurate recommendation mechanism, which limits the quality of user experience and social interaction, especially today when data information changes dramatically. The purpose of the present invention is to specifically target scenarios of big data and users' multiple interests and to provide a method and system for recommending users of a social network. In the recommendation of users of a social network, the influence of users' potential behaviors on the recommended users is taken into account, and a method and system for recommending users of a social network based on analysis and prediction of users' multiple behaviors and potential interests is proposed, aiming to accurately identify and recommend potential social relationships that users may be interested in through data modeling, thereby improving user experience and the activity of the social network platform.

[0005] The purpose of the present invention is mainly achieved through the following technical solutions: A method for recommending users on a social network, comprising the following steps:

[0006] Step S1, collect the behavior information data of the system registered users, the behavior information data includes user data, behavior data and interest point data, the behavior data includes social interaction data, the behavior information data is vectorized through the word vector model and four vector groups are output, the four vector groups are: user vector group, behavior vector group, user interest vector group and social interaction vector group.

[0007] Step S2: Analyze and learn the behavior information data using a deep neural network model to embed the user vector and the user interest vector, thereby completing the initialization embedding of the user vector and the user interest vector, and outputting the first behavior in which the embedding of the user vector and the user interest vector is initialized;

[0008] The behavior vectors in the behavior vector group are embedded into user vectors and user interest vectors through the graph convolution network of the deep neural network model, and the embedding learned in the behavior vector is used as the input feature of the subsequent behavior vector to capture the dependency of the behavior vectors in the embedding learning process and establish the cascade relationship of multiple behavior vectors of the user vector;

[0009] The cascade relationship of multiple behavior vectors is obtained by decoupling representation technology to obtain different factor vectors, and the influence values ​​of different factor vectors on different user preference scores are modeled. The embedding of user vectors and user interest vectors in all behavior vectors is used as the personalized features of user vectors and user interest vectors. User vectors and interest feature vectors are extracted from the current behavior vector as meta-units. The meta-units are calculated through a two-layer feedforward neural network to form a meta-network. The meta-network generates a transformation matrix to represent the personalized feature transformation of different user vectors and user interest vectors in different behavior vectors.

[0010] The influence values ​​of different factor vectors on different user preference scores are the preference scores of user vectors to user interest vectors calculated by the aggregation embedding method. The preference scores are sorted to obtain user interest tags as user vector portraits.

[0011] Step S3: Based on the user vector portrait and social interaction vector, the similarity matrix and hierarchical clustering method are used in turn to perform two-stage screening of potential friends, match possible social relationships between user vectors, and recommend potential friends with similar interests according to the set threshold number.

[0012] A further optimization solution is to vectorize the behavior information data through the word vector model and output four sets of vector groups, including:

[0013] The user data is trained using the word vector model and the user vector is recorded as u, u∈U=(u1,u2,…,u M ), M represents the total number of user vectors, U represents the user vector group, u M Represents the user vector of the Mth user.

[0014] The behavior data and interest point data are trained using the word vector model and recorded as G b , G b ∈G=(G 1 ,G 2 ,…,G B), G represents the behavior vector group, b represents the bth behavior in the vector group, that is, the current behavior vector is recorded as behavior vector b, B represents the total number of behavior vectors, and the trained user interest vector is recorded as i, i∈I=(i1,i2,...,i N ), I represents the user interest vector group, and N represents the user interest vector type.

[0015] Constructing a user-social trust matrix based on social interaction vector groups X uu' Represents the interaction between user vector u and user vector u'. Interaction between users is recorded as 1, and no interaction is recorded as 0.

[0016] The word vector model is a model that maps text to a high-dimensional vector space, so that short words in the text are represented in this vector space, and the semantic information of the words is encoded into vectors. Words with similar semantics are also close in the vector space. The basic idea is to capture the semantic relationship between words by learning the context of a word, so that these vectors can mathematically reflect the similarities and relationships between words.

[0017] A further optimization scheme is that the first line of embedding of the user vector and the user interest vector is initialized, and the specific output steps are as follows:

[0018] The user vector u and user interest vector i are initialized as and d represents the size of the embedding, Indicates the domain, Represents a domain of size d.

[0019] make and They are used to represent the embedding matrices of user vectors and user interest vectors, respectively. M and N are the total number of user vectors and the total number of user interest vectors, respectively.

[0020] Each user vector and user interest vector is uniquely represented by a vector by using ID U and ID I Represent the vector embeddings of all user vectors and user interest vectors, and They are user vectors u m and user interest vector i n Vector.

[0021] User vector u m and user interest vector i n The first row of the embedding is initialized to be a vector:

[0022] A further optimization scheme is that the specific steps for constructing the cascade relationship of multiple behavior vectors of the user vector are as follows:

[0023] The embedding of user vector and user interest vector in behavior vector b is recorded as and Perform graph convolution on the user vector-user interest vector interaction graph of behavior vector b to obtain

[0024]

[0025] and They represent the embedding of user vector u and user interest vector i under behavior vector b after being updated through the lth layer, is the normalization term, N u The set of user interest vectors interacting with user vector u, N i The set of user vectors that interact with user interest vector i.

[0026] The first embedding is recorded as l = 0, and there are L layers of embedding in total. After L layers of embedding, the embedding of each layer is fused to construct the final comprehensive embedding of the user vector and the user interest vector. The final comprehensive embedding is also the cascade relationship of multiple behavior vectors, which is expressed as: The user vector u and the user interest vector i learned under the behavior vector b are expressed as:

[0027]

[0028] Used to control aggregation, representing the weights of embeddings at layer l.

[0029] The further optimization scheme is that the steps of constructing personalized features are as follows:

[0030] The embedding obtained from the behavior vector b uses the decoupled representation technology to identify and model the impact values ​​of different factor vectors on different user preference scores. Because different types of behavior vectors reflect different aspects of user preferences, it is a comprehensive embedding of user vectors and user interest vectors, decoupling and representing each independent factor vector in various behavior vectors, ensuring that each factor vector remains independent. This separation allows more accurate modeling of user preference scores based on these individual factor vectors, and better mining of the potential interests of user vectors.

[0031] Using the decoupled representation technique, the embeddings of the user vector and user interest vector learned in each behavior vector are evenly divided into K independent blocks, that is, the factor vector is represented as:

[0032] and

[0033] They represent the kth block of user vector and user interest vector embedding respectively, d represents the embedding dimension, and K represents the total number of divided blocks.

[0034] Each block represents an independent factor vector to more effectively mine the diverse preferences of user vectors for user interest vectors in different behavior vectors. The embeddings of user vectors and user interest vectors in all behavior vectors are divided into K blocks and used as personalized features.

[0035] The further optimization scheme is that the personalized feature conversion steps are as follows:

[0036] The personalized feature mapping of the user vector and the user interest vector is to transition from the current behavior vector b to the next behavior vector, and extract the meta-units of the user vector and the interest feature vector from the behavior vector b.

[0037] K blocks are embedded one by one, and the definition of the unit cell in the bth block is as follows:

[0038] They represent the meta-units of the k-th block embedding of the user vector and the user interest vector in the behavior vector b, respectively; Represents the kth block embedding of the user vector and user interest vector learned in the behavior vector b.

[0039] They correspond to the aggregate embedding of the first-order potential friends around the user vector and user interest vector learned from the previous behavior vector, respectively.

[0040] Graph convolution takes the k-th block embedding of the user vector and user interest vector’s own nodes and the k-th block embedding of the first-order potential friends of the previous behavior vector as meta-units; it combines the features learned from the interaction in the current behavior vector with the features directly inherited from the previous behavior vector to capture the personalized features of the user vector and user interest vector.

[0041] The user vector and user interest vector personalized transformation matrix are generated using the meta-units collected in the meta-network; the meta-network is defined as follows:

[0042]

[0043] f(…) represents the meta-network, which uses a two-layer feedforward neural network. is a personalized transformation matrix and is embedded into the kth block of the corresponding user vector and user interest vector in the behavior vector b.

[0044] The further optimization solution is that the specific steps of constructing the user vector portrait are as follows:

[0045] The user interest vector i in the behavior vector b uses the attention mechanism to calculate the user preference score of the user vector u for the user interest vector i through the aggregation embedding method:

[0046]

[0047] Y ui Get the user preference score of user vector u for user interest vector i, represents the aggregation of the preference of user vector u for the kth factor vector, taking into account the preference degree in different behavior vectors, The preference score of the kth factor vector of the user interest vector is calculated using the information of all behavior vectors.

[0048] Finally, all K factor vectors are scored, sorted according to the scores to make user interest tags as user vector portraits, and recorded as user vector-user interest vector matrix

[0049] A further optimization plan is that the first stage of screening potential friends in the two-stage screening of potential friends specifically includes:

[0050] In the first stage, potential friends are screened using the user vector-user interest vector matrix and Pearson similarity calculation to calculate the similarity of the fusion attribute preferences of user vector u and user vector v. The similarities are sorted according to the calculated similarities to find the W user vectors with the highest similarity as potential friends.

[0051]

[0052] Credibility coefficient I u ∩I v represents the common interest item set of user vectors u and v, |I u ∩I v | represents the number of common rating items of two user vectors, |I u ∪I v | represents the number of all rating items of user vectors u and v, I s represents the common attribute set of non-commonly rated items of two user vectors, sim f (u,v) represents the similarity between user vectors u and v, R ui represents the preference of user vector u to user interest vector i, represents the mean score of user vector u, and user vector v represents the same, R vi represents the preference of user vector v to user interest vector i, Represents the mean rating of user vector v.

[0053] The further optimization plan is that the second stage of screening potential friends in the two-stage screening of potential friends specifically includes:

[0054] The user vector-user interest vector matrix uses a hierarchical clustering method to cluster the W potential friends screened in the second stage, and specifies the number of clusters as Y. After the clustering is completed, the potential friends in each category have similar interests and hobbies. From each category, select a potential friend with the highest similarity to the target user vector, and finally get Y potential friends with similar interests.

[0055] The social network user recommendation system of the present invention comprises a sequentially connected data collection and integration module, a user vector portrait construction module and a potential friend matching recommendation module;

[0056] The data collection and integration module includes a data collection unit and a word vector model data integration unit;

[0057] The data collection unit collects the behavior information data of the registered users of the system. The behavior information data includes user data, behavior data and point of interest data. The behavior data also includes social interaction data. The word vector model data integration unit vectorizes the collected behavior information data through the word vector model and outputs four vector groups. The four vector groups are: user vector group, behavior vector group, user interest vector group and social interaction vector group, and makes a user-social trust matrix based on the social interaction vector group;

[0058] The user vector portrait construction module consists of a nesting unit, a decoupled cascade unit, and a personalized feature conversion unit;

[0059] The nesting unit uses the user vector, user interest vector, behavior vector and user-social trust matrix, adopts a deep neural network model for analysis and learning, embeds the user vector and user interest vector, and outputs the first behavior vector of the embedding initialization of the user vector and user interest vector;

[0060] The decoupling cascade unit nests the first behavior vector that initializes the embedding of the user vector and the user interest vector multiple times, establishes a cascade relationship between multiple behavior vectors of the user vector, and models the influence values ​​of different factor vectors on different user preference scores through the decoupling representation technology;

[0061] The personalized feature conversion unit maps the personalized features of the user vector and the user interest vector to transition from the current behavior vector to the next behavior vector, and extracts the user vector and the interest feature vector from the current behavior vector as a meta-unit. The meta-unit is calculated through a two-layer feedforward neural network to form a meta-network. The meta-network realizes the feature conversion of different user vectors and user interest vectors in different behavior vectors. The preference score of the user vector to the user interest vector is calculated using the attention mechanism by using the aggregation embedding method, and the preference scores are sorted to obtain the user interest tags as the user vector portrait;

[0062] The potential friend matching recommendation module includes two stages of screening potential friends. The first stage screens potential friends based on the user interest tags and social interaction vectors of the user vector portrait, and uses the similarity matrix to calculate the user vectors with similar interests. The second stage screens potential friends, based on the first stage of screening potential friends, the most similar W matrices are judged as user vectors with similar interests, and the hierarchical clustering method is used to screen potential friends in two stages, match the possible social relationships between user vectors, and recommend Y potential friends with the most similar interests.

[0063] The working principle of this solution is as follows: According to the behavioral information data, the word vector model is used to convert it into the corresponding vector group, and the relationship between the user's behavior, potential behavior and interest is analyzed through deep learning through decoupling representation technology and attention mechanism. The user vector portrait is made using the graph convolutional network, and the user interest vector is marked. The potential friends are screened and recommended through two stages of Pearson similarity calculation and hierarchical clustering method. A social network user recommendation method and system based on the analysis and prediction of user vectors and multiple behavior vectors is proposed.

[0064] In summary, compared with the prior art, the present invention has the following beneficial effects: the social network user recommendation method and system of the present invention, using advanced technical means and algorithms, solves the accuracy and personalization problems existing in traditional social network recommendation systems, provides user vectors with a better social experience, and has significant market application potential and economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation on the embodiments of the present invention. In the drawings:

[0066] Figure 1 A diagram of a social network user recommendation system;

[0067] Figure 2 A step-by-step diagram of a recommendation method for social network users;

[0068] Figure 3 A diagram of the steps of data collection and integration in a social network user recommendation method;

[0069] Figure 4 A diagram showing the steps for constructing a user vector profile in a social network user recommendation method;

[0070] Figure 5 Graph of the decoupled cascade steps for constructing user vector profiles in a user recommendation method for social networks. DETAILED DESCRIPTION

[0071] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0072] Traditional social networking platforms rely on user vectors to add friends or follow accounts on their own, and lack an accurate recommendation mechanism, which limits the quality of user vector experience and social interaction vectors, especially today when data information changes dramatically; in view of this, the present invention provides the following embodiments to solve the above technical problems.

[0073] The present invention provides a scenario specifically for recommending potential friends on a social network platform, and proposes a social network user recommendation method and system based on user vector multi-behavior vector and interest analysis and prediction, which can be applied and transplanted to different social network platforms. The principle of the present invention is:

[0074] According to the behavioral information data, the word vector model is used to convert it into the corresponding vector group, and the deep neural network model is used for analysis and learning. The graph convolutional network is used to establish the cascade relationship of multiple behavior vectors of the user vector. The decoupling representation technology and attention mechanism are used for deep learning to analyze the relationship between the user's behavior, potential behavior and interest, and the impact value of different factor vectors on different user preference scores is modeled. The user vector is profiled using the graph convolutional network, and the user interest vector is marked. The potential friends are screened and recommended in two stages through Pearson similarity calculation and hierarchical clustering method. A social network user recommendation method and system based on the analysis and prediction of user vectors and multiple behavior vectors is proposed.

[0075] Example:

[0076] A social network user recommendation system, such as Figure 1 As shown, it includes: a sequentially connected data collection and integration module, a user vector portrait construction module and a potential friend matching recommendation module.

[0077] The data collection and integration module includes a data collection unit and a word vector model data integration unit;

[0078] The data collection unit collects the behavior information data of the registered users of the system. The behavior information data includes user data, behavior data and point of interest data. The behavior data also includes social interaction data. The word vector model data integration unit vectorizes the collected behavior information data through the word vector model and outputs four vector groups. The four vector groups are: user vector group, behavior vector group, user interest vector group and social interaction vector group, and makes a user-social trust matrix based on the social interaction vector group;

[0079] The user vector portrait construction module consists of a nesting unit, a decoupled cascade unit, and a personalized feature conversion unit;

[0080] The nesting unit uses the user vector, user interest vector, behavior vector and user-social trust matrix, adopts a deep neural network model for analysis and learning, embeds the user vector and user interest vector, and outputs the first behavior vector of the embedding initialization of the user vector and user interest vector;

[0081] The decoupling cascade unit nests the first behavior vector that initializes the embedding of the user vector and the user interest vector multiple times, establishes a cascade relationship between multiple behavior vectors of the user vector, and models the influence values ​​of different factor vectors on different user preference scores through the decoupling representation technology;

[0082] The personalized feature conversion unit maps the personalized features of the user vector and the user interest vector as a transition from the current behavior vector to the next behavior vector, and extracts the user vector and the interest feature vector from the current behavior vector as a meta-unit. The meta-unit is calculated through a two-layer feedforward neural network to form a meta-network. The meta-network realizes the feature conversion of different user vectors and user interest vectors in different behavior vectors. The preference score of the user vector to the user interest vector is calculated using the attention mechanism using the aggregation embedding method. The preference scores are sorted to obtain the user interest tags as the user vector portraits.

[0083] The potential friend matching recommendation module includes two stages of screening potential friends. The first stage screens potential friends based on the user interest tags and social interaction vectors of the user vector portrait, and uses the similarity matrix to calculate the user vectors with similar interests. The second stage screens potential friends based on the first stage of screening potential friends to determine the most similar W matrices as user vectors with similar interests, and uses the hierarchical clustering method to perform two-stage screening of potential friends, match possible social relationships between user vectors, and recommend Y potential friends with the most similar interests.

[0084] A social platform using the social network user recommendation system of this embodiment, and a specific social network user recommendation method are as follows: Figure 2 , including 3 steps:

[0085] Step S1, collect the behavior information data of the system registered users, the behavior information data includes user data, behavior data and interest point data, the behavior data includes social interaction data, the behavior information data is vectorized through the word vector model and four vector groups are output, the four vector groups are: user vector group, behavior vector group, user interest vector group and social interaction vector group.

[0086] Step S2: Analyze and learn the behavior information data using a deep neural network model to embed the user vector and the user interest vector, thereby completing the initialization embedding of the user vector and the user interest vector, and outputting the first behavior in which the embedding of the user vector and the user interest vector is initialized;

[0087] The behavior vectors in the behavior vector group are embedded into user vectors and user interest vectors through the graph convolution network of the deep neural network model, and the embedding learned in the behavior vector is used as the input feature of the subsequent behavior vector to capture the dependency of the behavior vectors in the embedding learning process and establish the cascade relationship of multiple behavior vectors of the user vector;

[0088] The cascade relationship of multiple behavior vectors is obtained by decoupling representation technology to obtain different factor vectors, and the influence values ​​of different factor vectors on different user preference scores are modeled. The embedding of user vectors and user interest vectors in all behavior vectors is used as the personalized features of user vectors and user interest vectors. User vectors and interest feature vectors are extracted from the current behavior vector as meta-units. The meta-units are calculated through a two-layer feedforward neural network to form a meta-network. The meta-network generates a transformation matrix to represent the personalized feature transformation of different user vectors and user interest vectors in different behavior vectors.

[0089] The influence values ​​of different factor vectors on different user preference scores are the preference scores of user vectors to user interest vectors calculated by the aggregation embedding method. The preference scores are sorted to obtain user interest tags as user vector portraits.

[0090] Step S3: Based on the user vector portrait and the social interaction vector, the similarity matrix and the hierarchical clustering method are used in turn to perform two-stage screening of potential friends, match the possible social relationships between the user vectors, and recommend potential friends with similar interests according to the set threshold number. In this embodiment, the user vector group is composed of multiple user vectors, the behavior vector group is composed of multiple behavior vectors, the user interest vector group is composed of multiple user interest vectors, and the social interaction vector group is composed of multiple social interaction vectors.

[0091] Wherein step S1 is as follows Figure 3 As shown, specifically including:

[0092] The collection includes user data, behavior data, and point of interest data, and social interaction data is treated as a separate set of data in the user's behavior data, because social interaction data can directly reflect the existing connections between users, such as jointly paying attention to a topic, and user A liked user B's comment.

[0093] User data: collect basic information of users, including age and gender.

[0094] Behavioral data: Data collection of users’ behavioral activities on social networking platforms, including content posted, likes, comments, shares, click behaviors, browsing history, and search queries in the fields of travel, sports, food, film and television entertainment, and learning.

[0095] Point of interest data: collect user interest tags, topics of interest, community information of participation, and content types or topics that users frequently visit.

[0096] Social interaction data: Obtain the interaction history between users and other users, such as the number of private messages, likes, and comments.

[0097] Converting user data, behavior data, point of interest data, and social interaction data into vector representation is the key to ensuring data quality and consistency.

[0098] The user data is trained using the word vector model and the user vector is recorded as u, u∈U=(u1,u2,...,u M ), M represents the total number of user vectors, U represents the user vector group, u M Represents the user vector of the Mth user.

[0099] The behavior data and interest point data are trained using the word vector model and recorded as G b , G b ∈G=(G 1 ,G 2 ,...,G B ), G represents the behavior vector group, b represents the bth behavior in the vector group, that is, the current behavior vector is recorded as behavior vector b, B represents the total number of behavior vectors, and the trained user interest vector is recorded as i, i∈I=(i1,i2,...,i N ), I represents the user interest vector group, and N represents the user interest vector type.

[0100] Constructing a user-social trust matrix based on social interaction vector groups X uu' Represents the interaction between user vector u and user vector u'. Interaction between users is recorded as 1, and no interaction is recorded as 0.

[0101] The word vector model is a model that maps text to a high-dimensional vector space, so that short words in the text are represented in this vector space, and the semantic information of the words is encoded into vectors. Words with similar semantics are also close in the vector space. The basic idea is to capture the semantic relationship between words by learning the context of a word, so that these vectors can mathematically reflect the similarities and relationships between words.

[0102] In this embodiment, the word vector model adopts the Chinese short text sentiment classification model proposed by Yang in 2021, which can combine the context to obtain the specific meaning of polysemous words in the text. First, the pre-trained user data, behavior data and interest point data are learned by using the ELMo model. The BiLSTM network structure in the ELMo model makes the word vector context-sensitive, and the expression of polysemous words is more accurate, which can effectively improve the accuracy of sentiment classification. The final output vector group is more in line with the various behavior vectors and language habits of interests in Chinese short texts.

[0103] In step S2, a graph convolutional network is used to improve the user vector portrait construction.

[0104] First, user vectors show different preferences for different factor vectors in different behavior vectors, which requires a method to effectively disentangle and independently evaluate the impact of these changing factor vectors on user vectors and user interest vectors, as well as to calculate the preferences of user vectors for different user interest vector factor vectors.

[0105] Then, when modeling, we first use the graph convolutional network to find the cascade relationship between behavior vectors to ensure that the information obtained from one behavior vector can be passed to the next behavior vector, that is, the embedding of one behavior vector is the input of the next behavior vector.

[0106] And within each behavior vector, the decoupled representation technique is used to separate various factor vectors entangled in the embedding of user vectors and user interest vectors. The user vector, user interest vector and user behavior vector are integrated to form a unique meta-network, which promotes personalized feature conversion between behavior vectors, thereby improving the accuracy of recommendations.

[0107] Finally, the attention mechanism is used to capture the subtle preferences of user vectors for different factor vectors in different behavior vectors, so that the potential interests of user vectors can be mined, and the potential interests and user interest vectors are linearly aggregated together, and the user interest vectors are sorted according to the aggregated scores as the user vector portrait.

[0108] Step S2 is as follows Figure 4 As shown, specifically including:

[0109] Embedding initialization: Initialize the embedding of user vectors and user interest vectors, and output the first line of the embedding of user vectors and user interest vectors to be initialized as vectors, laying the foundation for the learning process of the entire model.

[0110] The user vector u and user interest vector i are initialized as and d represents the size of the embedding, Indicates the domain, Represents a domain of size d.

[0111] make and They are used to represent the embedding matrices of user vectors and user interest vectors, respectively. M and N are the total number of user vectors and the total number of user interest vectors, respectively.

[0112] Each user vector and user interest vector is uniquely represented by a vector by using ID U and ID I Represent the vector embeddings of all user vectors and user interest vectors, and They are user vectors u m and user interest vector i n Vector.

[0113] User vector u m and user interest vector i n The first row of the embedding is initialized to be a vector:

[0114] Due to the cascading nature of behavior vectors in the model, where the embeddings learned in one behavior vector serve as input features for subsequent behavior vectors, an effective pre-training operation is crucial. Insufficient embeddings learned in previous behavior vectors may adversely affect the learning process of subsequent behavior vectors.

[0115] Decoupled Cascade: At the core of the model, a graph convolutional network is used to deeply learn the embedding of user vectors and user interest vectors in each behavior vector, and the cascade relationship between behavior vectors is modeled. The cascade relationship between behavior vectors is utilized to ensure that the information obtained from one behavior vector can be passed to the next behavior vector. The decoupled representation technology is used to identify the impact of different factor vectors on different user preference scores, and ensure that these different factor vectors are independent of each other. The user vector and interest feature vector are extracted from the behavior vector b as meta-units. The meta-unit is calculated through a two-layer feedforward neural network to obtain a meta-network, generate personalized conversion moments for user vectors and user interest vectors, and capture the personalized features of user vectors and user interest vectors.

[0116] It is specially designed to capture the detailed preferences of user vectors over user interest vectors and promote the personalized feature conversion of user vectors and user interest vectors in various behavior vectors.

[0117] The decoupled cascade first constructs the cascade relationship of multiple behavior vectors of the user vector.

[0118] The embedding of user vector and user interest vector in behavior vector b is recorded as and Perform graph convolution on the user vector-user interest vector interaction graph of behavior vector b to obtain

[0119]

[0120] and They represent the embedding of user vector u and user interest vector i under behavior vector b after being updated through the lth layer, is the normalization term, N u The set of user interest vectors interacting with user vector u, N i The set of user vectors that interact with user interest vector i.

[0121] The first embedding is recorded as l = 0, and there are L layers of embedding in total. After L layers of embedding, the embedding of each layer is fused to construct the final comprehensive embedding of the user vector and the user interest vector. The final comprehensive embedding is also the cascade relationship of multiple behavior vectors, which is expressed as: The user vector u and the user interest vector i learned under the behavior vector b are expressed as:

[0122]

[0123] Used to control aggregation, representing the weights of embeddings at layer l.

[0124] Then personalized feature construction is performed.

[0125] The embedding obtained from the behavior vector b is further used to identify and model the impact of different factor vectors on different user preference scores using the decoupled representation technique. Because different types of behavior vectors reflect different aspects of user preferences, the user vector and the user interest vector are integrated and decoupled into independent factor vectors in various behavior vectors, ensuring that each factor vector remains independent. This separation allows for more accurate modeling of user preference scores based on these individual factor vectors, and better mining of the potential interests of user vectors.

[0126] Using the decoupled representation technique, the embeddings of the user vector and user interest vector learned in each behavior vector are evenly divided into K independent blocks, that is, the factor vector is represented as:

[0127] and

[0128] They represent the kth block of user vector and user interest vector embedding respectively, d represents the embedding dimension, and K represents the total number of divided blocks.

[0129] Each independent block represents an independent factor vector. Due to the cascade relationship of behavior vectors, these factor vectors are independent of each other. This independence enables the model to more effectively mine the diverse preferences of user vectors for user interest vectors in different behavior vectors, and divide the embeddings of user vectors and user interest vectors in all behavior vectors into K blocks; blocks in corresponding positions in different behavior vectors represent the same factor vector.

[0130] The inherent structure of the cascade relationship is exploited to ensure that the K block embeddings of the user vector and the user interest vector in the first behavior vector remain independent, and then these embeddings are passed one by one to the subsequent behavior vectors through block feature transformation, maintaining independence and consistency of factor vectors between behavior vectors. The process of maintaining independent block embeddings is repeated in each subsequent behavior vector.

[0131] Finally, personalized feature transformation is performed.

[0132] The meta-network formed by promoting customized feature extraction between different behavior vectors can better capture the different preferences of user vectors and provide more personalized and accurate preference representation covering various behavior vectors.

[0133] Personalized feature mapping of user vector and user interest vector is crucial for transitioning from the current behavior vector b to the next behavior vector, extracting the meta-units of user vector and interest feature vector from the behavior vector b, and retaining the most personalized and significant features of user vector and user interest vector.

[0134] K blocks are embedded one by one, and the definition of the unit cell in the bth block is as follows:

[0135] They represent the meta-units of the k-th block embedding of the user vector and the user interest vector in the behavior vector b, respectively; Represents the kth block embedding of the user vector and user interest vector learned in the behavior vector b.

[0136] They correspond to the aggregate embedding of the first-order potential friends around the user vector and user interest vector learned from the previous behavior vector, respectively.

[0137] During graph convolution, the k-th block embeddings of the user vector and user interest vector’s own nodes and the k-th block embeddings of their previous behavior vector’s first-order potential friends are used as meta-units; this strategy combines the features learned from the interactions in the current behavior vector with the features directly inherited from the previous behavior vector, aiming to more accurately capture the personalized features of the user vector and user interest vector.

[0138] The user vector and user interest vector personalized transformation matrix are generated using the meta-units collected in the meta-network; the meta-network is defined as follows:

[0139]

[0140] f(…) represents the meta-network, which uses a two-layer feedforward neural network. is a personalized transformation matrix and is embedded in the kth block of the corresponding user vector and user interest vector in the behavior vector b; the main function of the meta-network is to generate these customized transformation matrices, which are specially designed to be consistent with the unique personality characteristics of each user vector and user interest vector, capturing the subtle user preference scores in each different behavior vector, and effectively promoting the personalized feature conversion of user vectors and user interest vectors between different behavior vectors.

[0141] Considering the preference and relationship transfer between different behavior vectors, more detailed user preference scores are revealed in multi-behavior data. The user vector may pay more attention to a specific factor vector in a behavior vector, which indicates that the feature input transferred from the previous behavior vector to the current behavior vector should retain more information about this specific factor vector. Therefore, the transfer between behavior vectors is inherently personalized. For user interest vectors, the factor vectors that attract user interest vectors vary in different behavior vectors, which requires that the user interest vector information transfer between behavior vectors is closer to the personalized preferences of the user vector in each specific behavior vector.

[0142] After the personalized features are converted, the meta-network is used to complete the user interest tags. The attention mechanism is used through the aggregation embedding method to calculate the preference of user vector u for user interest vector i, comprehensively mine the user interest vector, and label the user vector as the user vector portrait.

[0143] The user interest vector i in the behavior vector b uses the attention mechanism to calculate the user preference score of the user vector u for the user interest vector i through the aggregation embedding method:

[0144]

[0145] Y ui Get the user preference score of user vector u for user interest vector i, represents the aggregation of the preference of user vector u for the kth factor vector, taking into account the preference degree in different behavior vectors, The preference score of the kth factor vector of the user interest vector is calculated using the information of all behavior vectors.

[0146] Finally, all K factor vectors are scored, sorted according to the scores to make user interest tags as user vector portraits, and recorded as user vector-user interest vector matrix

[0147] Step S3, potential friend matching recommendation adopts two stages to screen potential friends: based on the user vector portrait, after completing the calculation of the user interest vector dissimilarity matrix in the first stage, the second stage of potential friend screening is performed based on this. The first stage adopts Pearson similarity calculation to obtain the similarity between two user vectors, specifically constructing a similarity matrix of the user interest vector between the two user vectors, and selecting W potential friends who are most similar to the target user vector. On the basis of the first stage, the second stage of hierarchical clustering screening is performed based on the processed W potential friends to obtain the final Y potential friends who are similar to the target user vector and diverse from each other.

[0148] In the first stage, potential friends are screened using the user vector-user interest vector matrix and Pearson similarity calculation to calculate the similarity of the fusion attribute preferences of user vector u and user vector v.

[0149]

[0150] Sort by the calculated similarity and find the W user vectors with the highest similarity as potential friends.

[0151] Credibility coefficient I u ∩I v represents the common interest item set of user vectors u and v, |I u ∩I v | represents the number of common rating items of two user vectors, |I u ∪I v | represents the number of all rating items of user vectors u and v, I s represents the common attribute set of non-commonly rated items of two user vectors, sim f (u,v) represents the similarity between user vectors u and v, R ui represents the preference of user vector u to user interest vector i, represents the mean score of user vector u, and user vector v represents the same, R vi represents the preference of user vector v to user interest vector i, Represents the mean rating of user vector v.

[0152] In the second stage, the hierarchical clustering method is used to screen potential friends. The user vector-user interest vector matrix uses the hierarchical clustering method to cluster the W potential friends screened in the second stage, and specify the number of clusters as Y. After the clustering is completed, the potential friends in each category have similar interests and hobbies, and the interests of potential friends in different categories are quite different. From each category, a potential friend with the highest similarity to the target user vector is selected, and finally Y potential friends with similar interests are obtained.

[0153] In hierarchical clustering, "finding potential friends" means finding the user vector with the smallest distance to the user vector in each merging process.

[0154] In the first stage, we filter the potential friends and find the W most similar potential friends. Based on the user vector-user interest vector matrix of the user vector portrait, we take the user vector as a data point to calculate the similarity or distance matrix:

[0155] First, the user vector-user interest vector matrix is ​​used to calculate the similarity or distance between user vectors, and the commonly used Manhattan distance method is used to calculate the Manhattan distance between user vectors and user interest vectors; each user vector is regarded as an independent class, and initially each class contains only one user vector; the distance matrix between all classes is calculated to construct a distance matrix, which is initially just the distance matrix between user vectors; then the merge is started, the minimum distance or maximum similarity in the distance matrix is ​​found, and the corresponding two classes are merged; the distance matrix is ​​updated to reflect the distance between the new class and other classes; the distance is updated, and the distance between the new class and other classes is updated using the mean link of the merge strategy; the mean link, the distance between the new class and other classes is the mean distance of all points in the new class and the mean distance of points in other classes; the two steps of updating the distance and the mean link are repeated until all user vectors are merged into one class, or the predetermined Y class is found.

[0156] After screening potential friends according to the second stage, relevant potential friends are recommended to promote more social interaction vectors and relationship building.

[0157] In this embodiment, the system also uses a stream processing system to collect and organize data before processing the data using a word vector model.

[0158] In this embodiment, the behavioral information data of the system registered users is collected periodically through the stream processing system. The stream processing system is a system that processes real-time data streams. The stream processing system can process and analyze data streams in real time or near real time. These data streams can be continuous data from various data sources such as sensors, logs, social media, and transaction records.

[0159] Among them, the key features of the stream processing system include real-time processing. The stream processing system can process continuously arriving data streams and quickly generate results, which is crucial for the real-time response required in this embodiment and the social media analysis of user vectors; the low latency of the stream processing system. One of the goals of the design of the stream processing system is to reduce processing delays as much as possible. The time from data to results should be as short as possible. In this embodiment, the profiling of user vectors also requires subsequent modeling based on the collected data in a timely manner, so the low latency of the stream processing system is also very important; the stream processing system is scalable. The stream processing system needs to process a large amount of data streams, so it usually has a high degree of scalability and can be horizontally expanded to process more data streams; the stream processing system is fault-tolerant. In order to ensure the reliability of the system, the stream processing system usually has a fault-tolerant mechanism that can handle node failure problems, so that the amount of abnormal behavior information data output is reduced.

[0160] In this embodiment, the data collection and processing of user behavior information generally includes the following steps:

[0161] First, data collection is performed: the data source collected by the stream processing system can collect data from a variety of sources, including application logs, sensors, social media platforms, and transaction systems; the data collection tools collected by the stream processing system usually use data collection tools or agents to collect data and push it into the stream processing system.

[0162] Then the data stream is organized into event streams or record streams, which contain information about user vector behavior data and point of interest data; the data format can be structured JSON, Avro format or semi-structured log files.

[0163] Next, data processing is carried out: real-time calculations are performed first. The stream processing system performs real-time calculations on the arriving data stream through aggregation, filtering, and windowing operations; and data conversion is performed on the real-time calculation results to convert the original data into useful information and extract the user's points of interest from the click stream.

[0164] Finally, data storage and analysis: For temporary storage, the processed data can be temporarily stored in memory for subsequent analysis, or stored in a database or data warehouse for persistent storage; and for real-time analysis, the system can analyze real-time data and generate real-time reports.

[0165] Finally, we get user data and behavior data, and through the analysis of behavior data, we can infer the user's interest point data.

[0166] The collected data are firstly,collected and integrated, then the collected data are cleaned and,preprocessed, and the collected user data are vectorized.

[0167] Before the word vector model processes the data, the user's behavior information data collected will be used to remove outliers and noise data, and to process missing values ​​and incomplete data.

[0168] Integrate information from different data sources into a unified dataset and transmit it to the word vector model.

[0169] In specific embodiment 1, Figure 1 The system also includes real-time recommendation services and user vector feedback mechanisms to update user data in a closed loop.

[0170] The real-time recommendation judgment module includes a judgment unit and a data processing unit. The judgment unit judges whether the user vector agrees or rejects the potential friend recommendation. After the user vector performs the operation, the data processing unit is triggered to send an action signal to the data collection and integration module to collect new behavioral information data; dynamically adjust the user vector portrait and the potential friend matching recommendation results to ensure the timeliness and accuracy of the recommendation.

[0171] The user vector feedback module collects user vector feedback data and passes the feedback data to the data collection and integration module. It optimizes the recommendation algorithm and improves the recommendation intelligence level and user vector satisfaction.

[0172] The real-time recommendation judgment module and the user vector feedback module allow the system to continuously update and optimize the user vector portrait; when the real-time recommendation judgment triggers the streaming data system to collect the user vector behavior data and the user vector feedback mechanism to collect the user vector feedback data in real time, the user vector portrait is continuously updated and optimized to reflect the user vector's latest interest and behavior vector patterns.

[0173] Update user vector portraits based on real-time data, continuously optimize models and algorithms, and ensure the timeliness and accuracy of user vector portraits.

[0174] Combined with user vector feedback data, adjust interest weights and recommendation strategies to improve user vector satisfaction and recommendation accuracy.

[0175] The specific method of updating user data through a closed loop of real-time recommendation service and user vector feedback mechanism:

[0176] Step S4: After recommending potential friends in step S3, determine whether the user vector of the recommended potential friend agrees or refuses to recommend the potential friend, trigger step S1 to collect new behavior information data for the user vector, and adjust the user vector portrait in real time.

[0177] Step S5: Setting a user vector feedback mechanism, setting a user vector feedback page in the system, collecting user vector feedback data, and transmitting the user vector feedback data back to step S1.

[0178] The user vector feedback mechanism in step S5 includes: collecting user vector feedback, transmitting feedback data and user vector communication.

[0179] First, collect user vector feedback including:

[0180] Real-time feedback collection: a user vector feedback page with a feedback button or form is set up on the platform user vector interaction interface and recommendation result page, so that the user vector can directly provide feedback on the similarity of interests of recommended potential friends and the user interest vector.

[0181] Implicit feedback collection: Through user vector behavior data, we can check the click-through rate, browsing time, and collection of recommended potential friends, analyze the user vector's preference for recommended content, and extract features of the recommended potential friends accepted by the user vector.

[0182] Text feedback collection: sentiment analysis and topic extraction are performed on the text feedback provided by the user vectors to gain a deeper understanding of the needs and feelings of the user vectors. The behavioral vectors are trained through the word vector model and transmitted to data collection and integration.

[0183] Then, the feedback data is transmitted to step S1, which is to transmit the data collected from the user vector feedback.

[0184] Finally, it includes user vector communication, which will promptly provide feedback to user vectors on potential friends recommended based on the latest collected user vector behavior data, remind users to update recommended potential friends, enhance user vectors' trust and satisfaction with the system, design appropriate feedback response mechanisms, automated responses or manual customer service support, and ensure that user vectors feel valued and listened to after providing feedback.

[0185] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for recommending users on a social network, characterized in that: include: Step S1, collecting the behavior information data of the registered users of the system, wherein the behavior information data includes user data, behavior data and point of interest data, wherein the behavior data includes social interaction data, and vectorizing the behavior information data through a word vector model to output four vector groups, wherein the four vector groups are: user vector group, behavior vector group, user interest vector group and social interaction vector group; Step S2: Analyze and learn the behavior information data using a deep neural network model to embed the user vector and the user interest vector, thereby completing the initialization embedding of the user vector and the user interest vector, and outputting the first behavior in which the embedding of the user vector and the user interest vector is initialized; The behavior vectors in the behavior vector group are embedded into user vectors and user interest vectors through the graph convolution network of the deep neural network model, and the embedding learned in the behavior vector is used as the input feature of the subsequent behavior vector to capture the dependency of the behavior vectors in the embedding learning process and establish the cascade relationship of multiple behavior vectors of the user vector; The cascade relationship of multiple behavior vectors is obtained by decoupling representation technology to obtain different factor vectors, and the influence values ​​of different factor vectors on different user preference scores are modeled. The embedding of user vectors and user interest vectors in all behavior vectors is used as the personalized features of user vectors and user interest vectors. User vectors and interest feature vectors are extracted from the current behavior vector as meta-units. The meta-units are calculated through a two-layer feedforward neural network to form a meta-network. The meta-network generates a transformation matrix to represent the personalized feature transformation of different user vectors and user interest vectors in different behavior vectors. The influence value of different factor vectors on different user preference scores is the preference score of the user vector to the user interest vector calculated by the aggregate embedding method. The preference scores are sorted to obtain the user interest tags as the user vector portrait; Step S3: Based on the user vector portrait and social interaction vector, the similarity matrix and hierarchical clustering method are used in turn to perform two-stage screening of potential friends, match possible social relationships between user vectors, and recommend potential friends with similar interests according to the set threshold number.

2. A method for recommending users on a social network according to claim 1, characterized in that: The behavior information data is vectorized through the word vector model and then four vector groups are output, specifically including: The user data is trained using the word vector model and the user vector is recorded as u, u∈U=(u1,u2,...,u M ), M represents the total number of user vectors, U represents the user vector group, u M represents the user vector of the Mth user; The behavior data and interest point data are trained using the word vector model and recorded as G b , G b ∈G=(G 1 ,G 2 ,...,G B ), G represents the behavior vector group, b represents the bth behavior in the vector group, that is, the current behavior vector is recorded as behavior vector b, B represents the total number of behavior vectors, and the trained user interest vector is recorded as i, i∈I=(i1,i2,...,i N ), I represents the user interest vector group, and N represents the user interest vector type; Constructing a user-social trust matrix based on social interaction vector groups X uu' Represents the interaction between user vector u and user vector u'. Interaction between users is recorded as 1, and no interaction is recorded as 0.

3. A social network user recommendation method according to claim 1 or 2, characterized in that: The first line where the embedding of the user vector and the user interest vector is initialized has the following specific output steps: The user vector u and user interest vector i are initialized as and d represents the size of the embedding, Indicates the domain, represents a domain of size d; make and The embedding matrices used to represent user vectors and user interest vectors respectively, M and N are the total number of user vectors and the total number of user interest vectors respectively; Each user vector and user interest vector is uniquely represented by a vector by using ID U and ID I Represent the vector embeddings of all user vectors and user interest vectors, and They are user vectors u m and user interest vector i n A vector of User vector u m and user interest vector i n The first row of the embedding is initialized as a vector:

4. A method for recommending users on a social network according to claim 3, characterized in that: The specific steps of constructing the cascade relationship of multiple behavior vectors of the user vector are as follows: The embedding of user vector and user interest vector in behavior vector b is recorded as and Perform graph convolution on the user vector-user interest vector interaction graph of behavior vector b to obtain and They represent the embedding of user vector u and user interest vector i under behavior vector b after being updated through the lth layer, is the normalization term, N u The set of user interest vectors interacting with user vector u, N i The set of user vectors that interact with user interest vector i; The first embedding is recorded as l=0, and there are L layers of embedding in total. After L layers of embedding, the embedding of each layer is fused to construct the final comprehensive embedding of the user vector and the user interest vector. The final comprehensive embedding is also a cascade relationship of multiple behavior vectors, which is expressed as: The user vector u and the user interest vector i learned under the behavior vector b are expressed as: Used to control aggregation, representing the weights of embeddings at layer l.

5. A method for recommending users on a social network according to claim 4, characterized in that: The steps of constructing the personalized features are as follows: The embedding obtained from the behavior vector b uses the decoupling representation technique to identify and model the impact of different factor vectors on different user preference scores, decoupling and representing the independent factor vectors in various behavior vectors to ensure that each factor vector remains independent; Using the decoupled representation technique, the embeddings of the user vector and user interest vector learned in each behavior vector are evenly divided into K independent blocks, that is, the factor vector is represented as: They represent the kth block of the user vector and user interest vector embedding, d represents the embedding dimension, and K represents the total number of blocks divided; Each block represents an independent factor vector to mine the diverse preferences of user vectors for user interest vectors in different behavior vectors, and the embeddings of user vectors and user interest vectors in all behavior vectors are divided into K blocks and used as personalized features.

6. A method for recommending users on a social network according to claim 5, characterized in that: The steps of personalized feature conversion are as follows: The personalized feature mapping of the user vector and the user interest vector is to transition from the current behavior vector b to the next behavior vector, and extract the meta-units of the user vector and the interest feature vector from the behavior vector b; K blocks are embedded one by one, and the definition of the unit cell in the bth block is as follows: They represent the meta-units of the k-th block embedding of the user vector and the user interest vector in the behavior vector b, respectively; represents the kth block embedding of the user vector and user interest vector learned in the behavior vector b; They correspond to the aggregate embedding of the first-order potential friends around the user vector and user interest vector learned from the previous behavior vector, respectively; The graph convolution uses the k-th block embedding of the user vector and the user interest vector’s own nodes and the k-th block embedding of the first-order potential friends of the previous behavior vector as meta-units; Combine the features learned from the interactions in the current behavior vector with the features directly inherited from the previous behavior vector to capture the personalized features of the user vector and the user interest vector; The user vector and user interest vector personalized transformation matrix are generated using the meta-units collected in the meta-network; the meta-network is defined as follows: f(…) represents the meta-network, which uses a two-layer feedforward neural network. is a personalized transformation matrix and is embedded into the kth block of the corresponding user vector and user interest vector in the behavior vector b.

7. A method for recommending users on a social network according to claim 6, characterized in that: The specific steps of constructing the user vector portrait are as follows: The user interest vector i in the behavior vector b uses the attention mechanism to calculate the user preference score of the user vector u for the user interest vector i through the aggregation embedding method: Y ui Get the user preference score of user vector u for user interest vector i, represents the aggregation of user vector u’s preference for the kth factor vector, Calculate the preference score of the kth factor vector of the user interest vector using the information of all behavior vectors; Finally, all K factor vectors are scored, sorted according to the scores to make user interest tags as user vector portraits, and recorded as user vector-user interest vector matrix 8. A method for recommending users on a social network according to claim 7, characterized in that: The first stage of screening potential friends in the two-stage screening of potential friends specifically includes: In the first stage, potential friends are screened by using the user vector-user interest vector matrix and Pearson similarity calculation to calculate the similarity of the fusion attribute preferences of user vector u and user vector v. The similarities are sorted according to the calculated similarities, and the W user vectors with the highest similarity are found as potential friends. Credibility coefficient I u ∩I v represents the common interest item set of user vectors u and v, |I u ∩I v | represents the number of common rating items of two user vectors, |I u ∪I v | represents the number of all rating items of user vectors u and v, I s represents the common attribute set of non-commonly rated items of two user vectors, sim f (u,v) represents the similarity between user vectors u and v, R ui represents the preference of user vector u to user interest vector i, represents the mean score of user vector u, and user vector v represents the same, R vi represents the preference of user vector v to user interest vector i, Represents the mean rating of user vector v.

9. A method for recommending users on a social network according to claim 8, characterized in that: The second stage of screening potential friends in the two-stage screening of potential friends specifically includes: The user vector-user interest vector matrix uses a hierarchical clustering method to cluster the W potential friends screened in the second stage, and specifies the number of clusters as Y. After the clustering is completed, the potential friends in each category have similar interests and hobbies. A potential friend with the highest similarity to the target user vector is selected from each category, and finally Y potential friends with similar interests are obtained.

10. A social network user recommendation system, characterized in that: It includes a sequentially connected data collection and integration module, a user vector portrait construction module, and a potential friend matching recommendation module; The data collection and integration module includes a data collection unit and a word vector model data integration unit; The data collection unit collects the behavior information data of the registered users of the system, the behavior information data includes user data, behavior data and interest point data, and the behavior data also includes social interaction data. The word vector model data integration unit vectorizes the collected behavior information data through the word vector model and outputs four vector groups, which are: user vector group, behavior vector group, user interest vector group and social interaction vector group, and makes a user-social trust matrix according to the social interaction vector group; User vector portrait building module, including nesting unit, decoupled cascade unit and personalized feature conversion unit; The nesting unit uses the user vector, the user interest vector, the behavior vector and the user-social trust matrix, adopts a deep neural network model for analysis and learning, embeds the user vector and the user interest vector, and outputs a first behavior vector of the embedding initialization of the user vector and the user interest vector; The decoupling cascade unit nests the first behavior vector that initializes the embedding of the user vector and the user interest vector multiple times, establishes a cascade relationship of multiple behavior vectors of the user vector, and models the influence values ​​of different factor vectors on different user preference scores through the decoupling representation technology; The personalized feature conversion unit maps the personalized features of the user vector and the user interest vector to transition from the current behavior vector to the next behavior vector, and extracts the user vector and the interest feature vector from the current behavior vector as a meta-unit, the meta-unit is calculated through a two-layer feedforward neural network to form a meta-network, the meta-network realizes the feature conversion of different user vectors and user interest vectors in different behavior vectors, adopts the aggregation embedding method to calculate the preference score of the user vector to the user interest vector by using the attention mechanism, and sorts the preference scores to obtain the user interest tags as the user vector portrait; The potential friend matching recommendation module includes two stages of screening potential friends. The first stage of screening potential friends is based on the user interest tags and social interaction vectors of the user vector portrait, and the user vectors with similar interests are calculated using a similarity matrix. The second stage of screening potential friends is based on the first stage of screening potential friends to determine the most similar W matrices as user vectors with similar interests, and the hierarchical clustering method is used to perform two-stage screening of potential friends, match possible social relationships between user vectors, and recommend Y potential friends with the most similar interests.

Images