CN116842277A - Social recommendation method based on cross-topic comparison learning - Google Patents
Social recommendation method based on cross-topic comparison learning Download PDFInfo
- Publication number
- CN116842277A CN116842277A CN202310849422.5A CN202310849422A CN116842277A CN 116842277 A CN116842277 A CN 116842277A CN 202310849422 A CN202310849422 A CN 202310849422A CN 116842277 A CN116842277 A CN 116842277A
- Authority
- CN
- China
- Prior art keywords
- user
- topic
- social
- embedding
- under
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 54
- 230000006870 function Effects 0.000 claims abstract description 67
- 238000010586 diagram Methods 0.000 claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 14
- 230000003993 interaction Effects 0.000 claims abstract description 13
- 239000011159 matrix material Substances 0.000 claims description 34
- 230000006399 behavior Effects 0.000 claims description 19
- 238000005070 sampling Methods 0.000 claims description 17
- 239000013598 vector Substances 0.000 claims description 13
- 230000007246 mechanism Effects 0.000 claims description 11
- 230000002776 aggregation Effects 0.000 claims description 10
- 238000004220 aggregation Methods 0.000 claims description 10
- 230000003542 behavioural effect Effects 0.000 claims description 10
- 238000004364 calculation method Methods 0.000 claims description 5
- 230000004931 aggregating effect Effects 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 230000004069 differentiation Effects 0.000 claims description 3
- 238000005295 random walk Methods 0.000 claims description 3
- 238000011524 similarity measure Methods 0.000 claims description 3
- 230000008685 targeting Effects 0.000 claims description 3
- 230000008092 positive effect Effects 0.000 claims 1
- 230000000644 propagated effect Effects 0.000 abstract 1
- 239000002775 capsule Substances 0.000 description 19
- 238000010276 construction Methods 0.000 description 4
- 230000009466 transformation Effects 0.000 description 4
- 238000001914 filtration Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000008878 coupling Effects 0.000 description 2
- 238000010168 coupling process Methods 0.000 description 2
- 238000005859 coupling reaction Methods 0.000 description 2
- 238000005065 mining Methods 0.000 description 2
- 239000000203 mixture Substances 0.000 description 2
- 230000006855 networking Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000001125 extrusion Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/906—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
Abstract
The invention relates to a social recommendation method based on cross-topic comparison learning, which comprises the following steps: acquiring a user article interaction diagram in a recommendation system; performing article representation learning, and performing article clustering and dividing subjects; constructing social relations based on topics by cross-topic comparison; based on balance theory, the information is propagated in the subject based on SGCN; constructing an objective function of a self-supervision task under different topics according to an extended structure balance theory; adopting Bayesian personalized ranking loss in a recommendation task, and determining a total loss function of a recommendation system by combining a cross-theme comparison learning objective function and a self-supervision task objective function based on a balance theory; and training a recommendation system based on the total loss function to obtain a recommendation result considering the user difference. Compared with the prior art, the method and the device can be used for modeling the interest preference of the user facing different topics by constructing the high-order social relationship under different topics, so that the problem of conflict in constructing the social relationship based on the symbol network is relieved.
Description
Technical Field
The invention relates to the field of social networks, in particular to a social recommendation method based on cross-topic comparison learning.
Background
With the development of information technology, information in a network grows exponentially, and it becomes extremely difficult to process and select information required, in which case a recommendation system becomes an important method to solve the problem. The recommendation system models the user's historical behavioral data and the content of the item itself to recommend information to the user that they need and are interested in. Currently, the mainstream recommendation algorithms are generally classified into collaborative filtering-based methods, content-based methods and hybrid methods. Collaborative filtering is one of the most popular techniques for creating a recommendation system that can learn the interests of a user through their interaction history with items. Conventional collaborative filtering methods based on users utilize scores and preferences of similar users to generate recommendations for other users, but in real-world environments, the scoring matrix of users is usually very sparse, so there are not enough common scoring items to calculate the similarity of users, resulting in often poor recommender system performance.
Because low cost establishment of connections in an online social network may result in one person having too many friends in the network world, such that the social network is made up of valuable friends, occasional friends and event friends, users are not necessarily all similar, social relationships that mix useful and noisy connections may introduce negative information into the recommender system, and thus indiscriminately using all social relationships of the user may perform worse than traditional recommender systems.
Disclosure of Invention
The invention aims to provide a social recommendation method based on cross-topic comparison learning, which aims to relieve possible conflicts of social relations based on a symbol network, introduces multiple topics into a construction method of the symbol network, divides user interests into different topics and further relieves social noise of users; meanwhile, in order to distinguish the user interests under each topic as far as possible, a cross-topic comparison learning mechanism is adopted to maximize the user interest difference under different topics, and the user interest difference is integrated into a recommendation framework.
The aim of the invention can be achieved by the following technical scheme:
a social recommendation method based on cross-topic comparison learning comprises the following steps:
step 1) obtaining a user article interaction diagram in a recommendation system;
step 2) carrying out article representation learning based on a user article interaction diagram, clustering articles and dividing topics;
step 3) constructing social relations based on topics through topic comparison:
step 31) constructing a behavior social matrix of the user under different topics;
step 32) constructing an objective function of cross-topic comparison learning by adopting an InfoNCE method;
step 4) on the basis of SGCN information propagation in the subject:
step 41) calculating a user set of balanced and unbalanced paths from a user perspective;
step 42) obtaining user embedment of user balance or unbalance through information aggregation by utilizing a balance theory;
step 43) aggregating user interests under each topic by means of an attention mechanism;
step 5) self-supervision method based on extended balance theory:
step 51) following the extended structure balance theory, constructing an objective function of the self-supervision task under different topics;
step 52) adopting Bayesian personalized ranking loss in the recommendation task, and determining the total loss function of the recommendation system by combining a cross-topic comparison learning objective function and a self-supervision task objective function based on a balance theory;
step 53) training the recommendation system based on the total loss function to obtain recommendation results considering the user differences.
Said step 2) comprises the steps of:
step 21), carrying out node sequence sampling of a user article interaction diagram by adopting a biased random walk method, after obtaining a sampling sequence, carrying out article representation learning in a view by adopting a Skip-gram model and a negative sampling method, and training to obtain article embedding;
step 22) embedding the trained articles into clusters to obtain corresponding user interest topics and initial embedding of users in each topic.
The article representation learning is carried out in the view by adopting a Skip-gram model and a negative sampling method, and the training to obtain the article embedding is specifically as follows:
the Skip-gram model gets node representation by targeting the average log probability of the maximized sequence:
wherein c is the context window of the sequence, T is the sequence length, i t Represents the t-th item in the sequence, p (i) being the most important in Skip-gram k+j |i t ) The softmax function was used as:
wherein ,ei and e′i Is a vector representation of the input and output of the item node, N is the total item number;
because of the high time complexity of computing the denominator of the above equation, using a negative sampling method to approximately maximize the logarithmic probability of the softmax function, the above equation is expressed as:
where k is a negative sample derived from a negative sampling method, where the negative sample can be derived from a negative sample set at the time of construction of the view of the item; sigma is a sigmoid function;
the embedding of the item is trained according to the Skip-gram model.
The step 31) specifically comprises the following steps:
the social relationships of users are divided into two categories: original social network G s And behavioural social network G b A for adjacency matrix of original social network s The social network of user behavior refers to an implicit social network in which no social relationship exists among users but historical behaviors are generated on the same article, and the social network of user behavior is called a positive social network diagramAnd negative behavioral social network diagram->Adjacency matrix of the active social network graph is +.>Adjacency matrix of passive behavior social network graph isIts calculation under the topic t is as follows:
wherein ,the positive and negative social behavioural matrices under the topic t respectively,a user positive scoring matrix and a user negative scoring matrix, respectively, for a topic t, wherein if the user's score is for the topic tr > 3, his score is considered positive and vice versa. Merging social relations, which are regarded as positive by the original social networks, into positive behavior social relations, wherein the positive social relations are adjacent to a matrix under a theme tNegative social relation adjacency matrix under topic t>
The step 32) specifically includes the following steps:
determining active neighbor embedding of the user u under the topic t, wherein the neighbor embedding is expressed as:
wherein , and />Embedding for the active neighbor and the active neighbor of user u under topic t, respectively, < >>Is an adjacency matrix->Row vector centered on user u +.>Measuring the neighbor connection number of the user u under the topic t;
given a user u and a topic t, the objective function obtained by using info NCE is:
wherein ,vut For user u user interest under topic t, t ′ For other subjects, τ is a superparameter of the objective function, used to control the degree of differentiation of the model to the negative samples;
determining an objective function of cross-topic comparison learning based on the objective function of the user under the topic:
where m is the number of users and K is the number of topics.
The user set that computes balanced and unbalanced paths from a user perspective is defined as:
wherein ,representing a positive friend, ++1 for user i when the path length under topic t is l->Representing the passive neighbors of user i with path length l+1 under topic t, u representing the user.
The user embedding for obtaining user balance or unbalance through information aggregation is as follows:
wherein , and />Representing user u user positive and negative embedment under theme t->Andas parameters to be trained, namely: each layer of information aggregation under the theme, the positive embedding is composed of the positive embedding of the positive neighbors and the negative embedding of the negative neighbors, wherein the positive embedding is composed of the positive interests of the last layer of the user;
in particular, when l=1, the user active embedding is aggregated only by the embedding of its active neighbors, while the user passive embedding is aggregated only by the passive neighbor embedding of the user:
after the positive embedding and the negative embedding of the user under different topics are obtained, the positive embedding and the negative embedding of the user are aggregated in the topics by utilizing a neural network, and the user embedding in the topics is obtained:
wherein MLP stands for multi-layer perceptron.
The aggregation of the user interests under each theme through the attention mechanism is specifically to coherently aggregate the user embedments of different themes through the attention mechanism, so as to obtain complete user interest embedments:
wherein ,vu For complete user embedding, K is the number of topics, v ut For user u user interest under topic t, α t Attention coefficient for subject t:
wherein a and W T Is a parameter requiring training. The user embedding of different topics is coherently aggregated through the attention mechanism, so that complete user interest embedding is obtained.
And designing an objective function of a self-supervision task according to the extended structure balance theory, so as to correctly aggregate the active social embedding and the passive social embedding, and mining semantic information in the active social embedding and the passive social embedding. Within a topic t, the objective function of the self-supervising task is:
wherein , and />Representing the active and passive neighbors of the user within the topic t, v ut Embedded for user u's interest under topic t, < ->Is a similarity measure function.
The total loss function of the recommendation system is as follows:
wherein ,λc and λb In order to control the super-parameters of the auxiliary task effect,for the objective function of self-supervision task, +.>For the objective function of cross-topic contrast learning, +.>For Bayesian Personalized Ranking (BPR) penalty:
wherein phi represents the parameter of CTCL,for the predictive score of user u for item i, σ is a sigmoid function, λ r Is a regularization parameter.
Compared with the prior art, the invention has the following beneficial effects:
(1) In order to alleviate the possible conflict of social relations based on the symbol network, the method introduces multiple topics into the construction method of the symbol network, and divides the interests of the user into different topics, so that the social relations of the user are constructed under the different interests of the user, and the social noise of the user is further relieved.
(2) The invention distinguishes the user interests under each theme as far as possible, adopts a cross-theme comparison learning mechanism to maximize the user interest difference under different themes, integrates the user interest difference into a recommendation frame, and improves the recommendation performance.
Drawings
FIG. 1 is a schematic flow chart of the method of the present invention;
FIG. 2 is a flow chart of an embodiment of the present invention;
FIG. 3 is a schematic diagram of a dynamic routing method according to an embodiment of the present invention;
fig. 4 is a balanced or unbalanced path diagram of an SGCN in an embodiment of the present invention.
Detailed Description
The invention will now be described in detail with reference to the drawings and specific examples. The present embodiment is implemented on the premise of the technical scheme of the present invention, and a detailed implementation manner and a specific operation process are given, but the protection scope of the present invention is not limited to the following examples.
The embodiment first defines a social network:
definition 1: social network diagram. Social networking graph refers to a graph structure that consists of social connections between people or organizations. In a social networking graph, nodes represent people or organizations, and edges represent social connections between them.
Definition 2: equilibrium theory. In the triangular motif, equilibrium theory can be described as: four rules of "friend of friend is friend", "enemy of enemy is friend", "enemy of friend is enemy" and "enemy of enemy is enemy".
Definition 3: a user set U. A collection of users in the recommendation system.
Definition 4: item set I. The collection of items in the recommendation system.
Definition 5: and (5) a scoring matrix R. Description of user interactions with items in a recommendation system.
Definition 6: user friend set N t (u). User u's set of friends under topic t.
Definition 7: the adjacency matrix a. User hair adjacency matrix in social network, a ij =1 denotes that user i trusts user j in the social network, i.e. user j is a friend of user i.
Definition 8: the user embeds V. A low-dimensional vector representation of user features.
Definition 9: the article is embedded in E. A low-dimensional vector representation of the item features.
As shown in fig. 1, the embodiment provides a social recommendation method based on cross-topic comparison, and the flow architecture of the social recommendation method is shown in fig. 2, and includes the following steps:
step 1) obtaining a user article interaction diagram in a recommendation system.
And 2) carrying out item representation learning based on the user item interaction diagram, clustering the items and dividing the topics.
Step 21) carrying out node sequence sampling of the user article interaction diagram by adopting a biased random walk method, after obtaining a sampling sequence, carrying out article representation learning in the view by adopting a Skip-gram model and a negative sampling method, and training to obtain article embedding.
The Skip-gram model gets node representation by targeting the average log probability of the maximized sequence:
wherein c is the context window of the sequence, T is the sequence length, i t Represents the t-th item in the sequence, p (i) being the most important in Skip-gram k+j |i t ) The softmax function was used as:
wherein ,ei and ei ′ Is a vector representation of the input and output of the item node, N is the total item number;
because of the high time complexity of computing the denominator of the above equation, using a negative sampling method to approximately maximize the logarithmic probability of the softmax function, the above equation is expressed as:
where k is a negative sample derived from a negative sampling method, where the negative sample can be derived from a negative sample set at the time of construction of the view of the item; sigma is a sigmoid function;
the embedding of the item is trained according to the Skip-gram model.
Step 22) embedding the trained articles into clusters to obtain corresponding user interest topics and initial embedding of users in each topic.
Specifically, two-layer capsule is designed, as shown in FIG. 3, the first layer capsule is embedded by the user behavior sequence, wherein e i Embedding the ith article which represents the user interaction, and obtaining a new user behavior example e by the first layer of capsules through different linear mapping matrixes ik Second layer capsule I k Can be regarded as a user interest capsule. The dynamic routing method is to calculate the value of the second layer capsule by using the first layer capsule and the initial routing logic, and then dynamically update the routing logic according to the correlation between the first layer capsule and the second layer capsule, and complete the extraction of multiple interests of the user through continuous iteration. First, the first layer capsule e i The object of the linear transformation is to observe the object i from different angles, the observation vector e being obtained ik The calculation is as follows:
e ik ＝W k e i
wherein Wk ∈R d×d For interest sharing a transformation matrix, each interest capsule k employs a different transformation matrix W k In order to reduce the complexity of the model and make the plurality of initial interests of the user different from each other, W of the present embodiment k Is normally distributedIs included in the random matrix extracted from the matrix.
For each iteration, interest capsule I k Calculated as a weighted sum of all observation vectors:
in order for the length of the output vector of the capsule to represent the probability of the entity represented by the capsule in the current input, capsule I is of interest K Compression is performed using a squaring function (extrusion function):
wherein c is as defined above ik For observing vector e ik Capsule of interest S k Using softmax and routing logic b ik To calculate the coupling coefficient c ik ：
wherein bik Is dynamic routing logic, the value of which is defined by the compressed interest capsule S k And the prediction vector e ik The inner product dynamics of (2) is:
b ik ＝S k T e ik
the dynamic routing method first routes the logic b ik Initialized to 0, and then the coupling coefficient c is calculated according to the routing logic ik In order to ensure that the initial interests are different, the first layer of capsules are subjected to linear transformation, and then the interest capsules are obtained through iterative calculation. The dynamic routing process proposed herein iterates three times to get a good convergence, eventually resulting in the user's embedded interest under one topic denoted v uk ＝S k 。
In another embodiment, the training-derived items are embedded into clusters using K-means++, gaussian Mixture Model (GMM), or the like, in addition to the dynamic routing method mentioned above.
Step 3) constructing social relations based on topics through topic comparison:
step 31) constructing a behavior social matrix of the user under different topics.
The social relationships of users are divided into two categories: original social network G s And behavioural social network G b A for adjacency matrix of original social network s The social network of user behavior refers to an implicit social network in which no social relationship exists among users but historical behaviors are generated on the same article, and the social network of user behavior is called a positive social network diagramAnd negative behavioral social network diagram->Adjacency matrix of the active social network graph is +.>Adjacency matrix of passive behavior social network graph isIts calculation under the topic t is as follows:
wherein ,the positive and negative social behavioural matrices under the topic t respectively,a user positive scoring matrix and a user negative scoring matrix under a topic t, respectively, wherein if the user's score r > 3 under the topic t, his score is considered positive and vice versa. Merging social relations, which are regarded as positive by the original social networks, into positive behavior social relations, wherein the positive social relations are adjacent to a matrix under a theme tNegative social relation adjacency matrix under topic t>
Step 32) adopting an InfoNCE method to construct an objective function of cross-topic comparison learning.
First, determine the active neighbor embedding of user u under topic t, the neighbor embedding is expressed as:
wherein , and />Embedding for the active neighbor and the active neighbor of user u under topic t, respectively, < >>Is an adjacency matrix->Row vector centered on user u +.>Measuring the neighbor connection number of the user u under the topic t; in this way, active neighbor embedding of the user under different topics is obtained.
Given a user u and a topic t, the objective function obtained by using info NCE is:
wherein ,vut For user u user interest under topic t, t ′ For other subjects, τ is a superparameter of the objective function, used to control the degree of differentiation of the model to the negative samples;
determining an objective function of cross-topic comparison learning based on the objective function of the user under the topic:
where m is the number of users and K is the number of topics.
Step 4) on the basis of SGCN information propagation in the subject:
step 41) calculates a user set of balanced and unbalanced paths from a user perspective.
wherein ,representing a positive friend, ++1 for user i when the path length under topic t is l->Representing the passive neighbors of user i with path length l+1 under topic t, u representing the user.
The balanced/unbalanced path for user u based on balance theory is shown in fig. 4.
Step 42) using the balance theory, user embeddings of user balance or unbalance are obtained by information aggregation.
wherein , and />Representing user u user positive and negative embedment under theme t->Andis a parameter requiring training. It can be understood that each layer of information aggregation under the subject, the active embedding is composed of the active embedding of the active neighbor and the passive embedding of the passive neighbor, which are the most interesting of the last layer of the user;
in particular, when l=1, the user active embedding is aggregated only by the embedding of its active neighbors, while the user passive embedding is aggregated only by the passive neighbor embedding of the user:
after obtaining positive and negative user embeddings under different topics, firstly, aggregating the positive and negative user embeddings in the topics by using a neural network, thereby obtaining the user embeddings in the topics:
wherein MLP stands for multi-layer perceptron.
Step 43) aggregating user interests under various topics through an attention mechanism:
wherein ,vu For complete user embedding, K is the number of topics, v ut For user u user interest under topic t, α t Attention coefficient for subject t:
wherein a and W T Is a parameter requiring training. The user embedding of different topics is coherently aggregated through the attention mechanism, so that complete user interest embedding is obtained.
Step 5) self-supervision method based on extended balance theory:
step 51) following the extended structure balance theory, the objective function of the self-supervising task is built under different topics.
And designing an objective function of a self-supervision task according to the extended structure balance theory, so as to correctly aggregate the active social embedding and the passive social embedding, and mining semantic information in the active social embedding and the passive social embedding. Within a topic t, the objective function of the self-supervising task is:
wherein , and />Representing the active and passive neighbors of the user within the topic t, v ut Embedded for user u's interest under topic t, < ->Is a similarity measure function.
Step 52) adopting Bayesian personalized ranking loss in the recommendation task, and combining the objective function of cross-topic comparison learning and the self-supervision task objective function based on the balance theory to determine the total loss function of the recommendation system.
Bayesian Personalized Ranking (BPR) loss is expressed as:
wherein phi represents the parameter of CTCL,for the predictive score of user u for item i, σ is a sigmoid function, λ r Is a regularization parameter.
After setting the objective function of the main recommendation task, unifying the cross-topic comparison learning and the balance theory in the topic as auxiliary tasks into a recommendation frame, and obtaining the total loss function of the recommendation system as follows:
wherein ,λc and λb In order to control the super-parameters of the auxiliary task effect,for the objective function of self-supervision task, +.>For comparison of learned objective functions across topics. Based on cross-topic comparison learning and balance theory in the topic, rich information in the heterogeneous network is mined, so that better recommendation performance is obtained.
Step 53) training the recommendation system based on the total loss function to obtain recommendation results considering the user differences.
In summary, the user interests among different topics are mined by cross-topic comparison, the auxiliary self-supervision signals are constructed by adopting the balance theory of one of the basic social theories in the symbol network, and the rich information in the heterogeneous graph is fully mined, so that the semantic information of positive and negative links in the symbol network is mined. In particular, according to the extended structural balance theory, it is believed that under the same topic, the user's interests should be more similar to friends with positive links and less similar to nodes with negative links. Self-supervision signals are constructed based on balance theory under different topics and are integrated into a recommendation framework as auxiliary tasks, so that a good effect is achieved.
The foregoing describes in detail preferred embodiments of the present invention. It should be understood that numerous modifications and variations can be made in accordance with the concepts of the invention by one of ordinary skill in the art without undue burden. Therefore, all technical solutions which can be obtained by logic analysis, reasoning or limited experiments based on the prior art by a person skilled in the art according to the inventive concept shall be within the scope of protection defined by the claims.
Claims (10)
1. The social recommendation method based on cross-topic comparison learning is characterized by comprising the following steps of:
step 1) obtaining a user article interaction diagram in a recommendation system;
step 2) carrying out article representation learning based on a user article interaction diagram, clustering articles and dividing topics;
step 3) constructing social relations based on topics through topic comparison:
step 31) constructing a behavior social matrix of the user under different topics;
step 32) constructing an objective function of cross-topic comparison learning by adopting an InfoNCE method;
step 4) on the basis of SGCN information propagation in the subject:
step 41) calculating a user set of balanced and unbalanced paths from a user perspective;
step 42) obtaining user embedment of user balance or unbalance through information aggregation by utilizing a balance theory;
step 43) aggregating user interests under each topic by means of an attention mechanism;
step 5) self-supervision method based on extended balance theory:
step 51) following the extended structure balance theory, constructing an objective function of the self-supervision task under different topics;
step 52) adopting Bayesian personalized ranking loss in the recommendation task, and determining the total loss function of the recommendation system by combining a cross-topic comparison learning objective function and a self-supervision task objective function based on a balance theory;
step 53) training the recommendation system based on the total loss function to obtain recommendation results considering the user differences.
2. The social recommendation method based on cross-topic comparison learning according to claim 1, wherein the step 2) comprises the steps of:
step 21), carrying out node sequence sampling of a user article interaction diagram by adopting a biased random walk method, after obtaining a sampling sequence, carrying out article representation learning in a view by adopting a Skip-gram model and a negative sampling method, and training to obtain article embedding;
step 22) embedding the trained articles into clusters to obtain corresponding user interest topics and initial embedding of users in each topic.
3. The social recommendation method based on cross-topic comparison learning according to claim 2, wherein the learning of the item representation in the view by adopting the Skip-gram model and the negative sampling method is specifically that the training to obtain the item embedding is as follows:
the Skip-gram model gets node representation by targeting the average log probability of the maximized sequence:
wherein c is the context window of the sequence, T is the sequence length, i t Represents the t-th item in the sequence, p (i) being the most important in Skip-gram k+j |i t ) The softmax function was used as:
wherein ,ei and ei ′ Is a vector representation of the input and output of the item node, N is the total item number;
because of the high time complexity of computing the denominator of the above equation, using a negative sampling method to approximately maximize the logarithmic probability of the softmax function, the above equation is expressed as:
wherein k is a negative sample obtained according to a negative sampling method, and sigma is a sigmoid function;
the embedding of the item is trained according to the Skip-gram model.
4. The social recommendation method based on cross-topic comparison learning according to claim 1, wherein the step 31) specifically comprises:
the social relationships of users are divided into two categories: original social network G s And behavioural social network G b A for adjacency matrix of original social network s The social network of user behavior refers to an implicit social network in which no social relationship exists among users but historical behaviors are generated on the same article, and the social network of user behavior is called a positive social network diagramAnd negative behavioral social network diagram->Adjacency matrix of the active social network graph is +.>Adjacent to passive behavior social network graphThe grafting matrix is->Its calculation under the topic t is as follows:
wherein ,a positive and a negative social behavioural matrix under topic t, respectively, +.>Respectively a user positive scoring matrix and a user negative scoring matrix under the topic t, and integrating the social relationship which is regarded as positive by the original social network into the positive action social, wherein the positive social relationship is adjacent to the matrix under the topic tNegative social relation adjacency matrix under topic t>
5. The social recommendation method based on cross-topic comparison learning according to claim 1, wherein the step 32) specifically comprises the following steps:
determining active neighbor embedding of the user u under the topic t, wherein the neighbor embedding is expressed as:
wherein , and />Embedding for the active neighbor and the active neighbor of user u under topic t, respectively, < >>Is an adjacency matrixRow vector centered on user u +.>Measuring the neighbor connection number of the user u under the topic t;
given a user u and a topic t, the objective function obtained by using info NCE is:
wherein ,vut For user u user interest under topic t, t ′ For other subjects, τ is a superparameter of the objective function, used to control the degree of differentiation of the model to the negative samples;
determining an objective function of cross-topic comparison learning based on the objective function of the user under the topic:
where m is the number of users and K is the number of topics.
6. The social recommendation method based on cross-topic comparison learning of claim 1, wherein the user set that calculates balanced and unbalanced paths from the user's perspective is defined as:
wherein ,representing a positive friend, ++1 for user i when the path length under topic t is l->Representing the passive neighbors of user i with path length l+1 under topic t, u representing the user.
7. The social recommendation method based on cross-topic comparison learning according to claim 6, wherein the user obtaining user balance or unbalance through information aggregation is embedded as follows:
wherein , and />Representing user u user positive and negative embedment under theme t->Andas parameters to be trained, namely: each layer of information aggregation under the theme, the positive embedding is composed of the positive embedding of the positive neighbors and the negative embedding of the negative neighbors, wherein the positive embedding is composed of the positive interests of the last layer of the user;
when l=1, the user positive embedding is aggregated only by the embedding of its positive neighbors, while the user negative embedding is aggregated only by the negative neighbor embedding of the user:
after the positive embedding and the negative embedding of the user under different topics are obtained, the positive embedding and the negative embedding of the user are aggregated in the topics by utilizing a neural network, and the user embedding in the topics is obtained:
wherein MLP stands for multi-layer perceptron.
8. The social recommendation method based on cross-topic comparison learning according to claim 1, wherein the aggregation of the user interests under each topic by the attention mechanism is specifically that the user embedments of different topics are coherently aggregated by the attention mechanism to obtain a complete user interest embedment:
wherein ,vu For complete user embedding, K is the number of topics, v ut For user u user interest under topic t, α t Attention coefficient for subject t:
wherein a and W T Is a parameter requiring training.
9. The social recommendation method based on cross-topic comparison learning according to claim 1, wherein in a topic t, the objective function of the self-supervision task is:
wherein , and />Representing the active and passive neighbors of the user within the topic t, v ut Embedded for user u's interest under topic t, f (& gt): ->Is a similarity measure function.
10. The social recommendation method based on cross-topic comparison learning according to claim 1, wherein the total loss function of the recommendation system is:
wherein ,λc and λb In order to control the super-parameters of the auxiliary task effect,for the objective function of self-supervision task, +.>For the objective function of cross-topic contrast learning, +.>Personalized ranking penalty for bayes:
wherein phi represents the parameter of CTCL,for the predictive score of user u for item i, σ is a sigmoid function, λ r Is a regularization parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310849422.5A CN116842277A (en) | 2023-07-11 | 2023-07-11 | Social recommendation method based on cross-topic comparison learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310849422.5A CN116842277A (en) | 2023-07-11 | 2023-07-11 | Social recommendation method based on cross-topic comparison learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116842277A true CN116842277A (en) | 2023-10-03 |
Family
ID=88159823
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310849422.5A Pending CN116842277A (en) | 2023-07-11 | 2023-07-11 | Social recommendation method based on cross-topic comparison learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116842277A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117131282A (en) * | 2023-10-26 | 2023-11-28 | 江西财经大学 | Multi-view graph contrast learning recommendation method and system integrating layer attention mechanism |
-
2023
- 2023-07-11 CN CN202310849422.5A patent/CN116842277A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117131282A (en) * | 2023-10-26 | 2023-11-28 | 江西财经大学 | Multi-view graph contrast learning recommendation method and system integrating layer attention mechanism |
| CN117131282B (en) * | 2023-10-26 | 2024-01-05 | 江西财经大学 | Multi-view graph contrast learning recommendation method and system integrating layer attention mechanism |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112905900B (en) | Collaborative filtering recommendation method based on graph convolution attention mechanism | |
| CN111931062B (en) | Training method and related device of information recommendation model | |
| CN110717098B (en) | Meta-path-based context-aware user modeling method and sequence recommendation method | |
| CN111881350B (en) | Recommendation method and system based on mixed graph structured modeling | |
| CN111523047A (en) | Multi-relation collaborative filtering algorithm based on graph neural network | |
| CN110795641B (en) | Network rumor propagation control method based on representation learning | |
| CN110032679B (en) | Dynamic news recommendation method based on hierarchical attention network | |
| CN112861967A (en) | Social network abnormal user detection method and device based on heterogeneous graph neural network | |
| CN111143704B (en) | Online community friend recommendation method and system integrating user influence relationship | |
| CN110737730B (en) | User classification method, device, equipment and storage medium based on unsupervised learning | |
| Huo et al. | Link prediction with personalized social influence | |
| CN116842277A (en) | Social recommendation method based on cross-topic comparison learning | |
| CN111949885A (en) | Personalized recommendation method for scenic spots | |
| CN109933720B (en) | Dynamic recommendation method based on user interest adaptive evolution | |
| CN117216281A (en) | Knowledge graph-based user interest diffusion recommendation method and system | |
| CN114298783A (en) | Commodity recommendation method and system based on matrix decomposition and fusion of user social information | |
| Hawkins et al. | Generalizing meanings from partners to populations: Hierarchical inference supports convention formation on networks | |
| CN112364245B (en) | Top-K movie recommendation method based on heterogeneous information network embedding | |
| CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
| CN117251586A (en) | Multimedia resource recommendation method, device and storage medium | |
| KR20210079264A (en) | Methods and systems for defining emotional machines | |
| CN114936890A (en) | Counter-fact fairness recommendation method based on inverse tendency weighting method | |
| CN112288195A (en) | Method and device for predicting central user behavior based on friend behavior and electronic equipment | |
| Zhou et al. | Stance Inference in Twitter through Graph Convolutional Collaborative Filtering Networks with Minimal Supervision | |
| Özer et al. | A machine learning-based framework for predicting game server load |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |