CN117112914B - Group recommendation method based on graph convolution - Google Patents
Group recommendation method based on graph convolution Download PDFInfo
- Publication number
- CN117112914B CN117112914B CN202311374128.XA CN202311374128A CN117112914B CN 117112914 B CN117112914 B CN 117112914B CN 202311374128 A CN202311374128 A CN 202311374128A CN 117112914 B CN117112914 B CN 117112914B
- Authority
- CN
- China
- Prior art keywords
- group
- user
- article
- layer
- population
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 62
- 230000003993 interaction Effects 0.000 claims abstract description 57
- 238000012549 training Methods 0.000 claims abstract description 27
- 239000013598 vector Substances 0.000 claims description 35
- 230000002452 interceptive effect Effects 0.000 claims description 12
- 230000002776 aggregation Effects 0.000 claims description 11
- 238000004220 aggregation Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 11
- 238000010586 diagram Methods 0.000 claims description 9
- 238000006116 polymerization reaction Methods 0.000 claims description 8
- 238000011002 quantification Methods 0.000 claims description 8
- 238000004519 manufacturing process Methods 0.000 claims description 6
- 230000004927 fusion Effects 0.000 claims description 4
- 230000005484 gravity Effects 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000013473 artificial intelligence Methods 0.000 abstract description 2
- 230000000694 effects Effects 0.000 description 12
- 230000006870 function Effects 0.000 description 8
- 238000013528 artificial neural network Methods 0.000 description 7
- 238000009792 diffusion process Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 5
- 238000010606 normalization Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 235000010627 Phaseolus vulgaris Nutrition 0.000 description 1
- 244000046052 Phaseolus vulgaris Species 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000003796 beauty Effects 0.000 description 1
- 230000001413 cellular effect Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012163 sequencing technique Methods 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
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
Abstract
The invention relates to the field of artificial intelligence, in particular to a group recommendation method based on graph convolution. According to the method, based on an undirected graph representing the interaction relationship between a user and an article, the interaction data between the user and the article is learned through a constructed user recommendation model, and user preference embedding comprising user preference information and article feature embedding comprising article feature information are obtained. The user preference embedment of the group members is aggregated to obtain an initial embedment of the group. And calculating cosine similarity among the groups through the common group members to obtain preference similarity among the groups. And training the initial embedding of the groups, the preference similarity among the groups, the characteristic embedding of the articles and the undirected graph input constructed by the undirected graph representing the interaction relation between the groups and the articles, and recommending the articles for the groups to be recommended by using the trained group recommendation model. The method and the device enrich group preference information based on the preference similarity among groups, improve group recommendation accuracy, and simultaneously maintain recommendation efficiency.
Description
Technical Field
The invention relates to the field of artificial intelligence, in particular to a group recommendation method based on graph convolution.
Background
With the rapid development of internet technology in the past decades, various internet and multimedia application layers are endless, and massive multimedia data such as texts, audios and videos are explosively increased, so that the problem of serious information overload is brought, and people can hardly search in massive data to obtain the desired result. The recommended technology plays a key role in solving the problem of information overload and is widely adopted by many online services. For example, when shopping online, the recommendation technology recommends the user with the goods which the user prefers to purchase, and when listening to music, the recommendation technology recommends the user with the songs which the user prefers, so that a great deal of time is saved for the user in the selection process, and convenience is brought to the life of people.
Existing recommendation systems focus on personalized recommendations, and with the advent of online life service platforms with social properties, such as beauty groups, public critique, mare's cellular networks, people's willingness to participate in collective activities and trends have grown rapidly in recent years, for example, going out to eat with colleagues, going out to travel with family members, going out to see movies with friends, participating in social activities, etc. However, since the group is composed of different group members, the different group members have different preferences, and have different effects in the group decision process, such scenes are complex and various, as shown in fig. 1, each person has a need to watch the movie by himself, and may also have a need to watch the movie together by composing the group, how to combine the viewing history of the person with the viewing history of the group, thereby recommending contents that are satisfactory to the group members, which is a challenging problem at present.
Similar to conventional recommendations, the population and items may be mapped in vector space as one point of vector space, respectively, and the degree of preference of the population for the items may be calculated by calculating the distance between the two points (i.e., the similarity of the two points). The item refers to content to be recommended. Under the ideal state, the group can learn the positions of the group and the articles in the vector space by using the historical interaction data of the group and the articles as the users, so that the embedded representation of the group with preference information is obtained, and the ideal recommendation effect is obtained. However, in reality, the interactions of the population with the items are very sparse compared to the interactions of the user with the items, and the information of the population preferences that can be learned from the historical interaction data of the population with the items is very limited. Thus, one basic approach is to aggregate embedded representations of groups with preference information using embedded representations of group members with preference information that can be learned from historical interaction data of dense group members with items, enriching the embedded representations of groups. There are various ways of preference aggregation, and conventional heuristic strategies have average, minimum pain, maximum satisfaction, etc. With the development of deep learning technology, researchers have proposed to learn weights of group members in a group by using an attention neural network, so as to learn better to obtain the preference of the group.
However, the complex preference aggregation mechanism has small promotion amplitude on the recommendation effect of the recommendation model in most cases, but greatly increases the time for model training and prediction. When applied to a large-scale recommendation scenario, the time required for model training and prediction increases even more by thousands of times, resulting in a dramatic decrease in the efficiency of recommendation using the recommendation model. At the same time, most researchers ignore the gain effect of higher-order interactions between groups on group recommendations, higher-order interactions refer to other interactions beyond simple user interactions with items or group interactions with items. Similar to users, the influence among the groups is similar to the interaction diffusion of the groups and the articles, the groups which interact the same articles together have the similarity, the relation of the higher-order interaction is modeled into a recommendation algorithm, deep preference information of the groups can be better learned, and some groups have the same group members, as shown in fig. 2, the similarity among different groups can be calculated through the common group members, and the groups with high similarity are regarded as having implicit interaction, so that the preference information of the groups is enriched.
Disclosure of Invention
In order to solve the problems, the invention provides a group recommendation method based on graph convolution.
The method comprises the following steps of:
step one, preparing user and article interaction data and group and article interaction data as training data, wherein the training data comprises usersArticle->Group->And (3) with the crowd->The method comprises the steps of carrying out a first treatment on the surface of the Use of user item relationship diagram->Representing the interaction relationship between the user and the article; use of group article relationship diagram->Representing interaction relationship between the group and the article;
step two, training data and a user article relation diagramInputting a user recommendation model convolution layer to obtain a user +.>User-embedded representation +.>And articles->Item embedded representation->;
Step three, according to the userUser-embedded representation +.>And articles->Item embedded representation->Quantification of user +.>For articles->Is +.>;
Step four, defining a loss function based on user personalized recommendationPre-training a user recommendation model to obtain user +.>Is embedded with user preferences->And articles->Is embedded with item characteristics->;
Step five, using an aggregation method to group the populationUser preferences of all group members of (a) are embedded and aggregated into +.>Population initial embedding->;
Step six, calculating the population through the common population membersAnd (3) with the crowd->Cosine similarity between them gives the population +.>And (3) with the crowd->Preference similarity between->;
Step seven, the group is formedPopulation initial embedding->Article->Is embedded with item characteristics->Group->And (3) with the crowd->Preference similarity between->And group article relationship diagram->Inputting a group recommendation model convolution layer to obtain a group +.>Group embedded representation->And articles->Item-embedding vector +.>;
Step eight, according to the groupGroup embedded representation->And articles->Item-embedding vector +.>Quantification of the population by dot product calculation>For articles->Is +.>;
Step nine, defining a loss function based on group recommendationTraining the group recommendation model to obtain a trained group recommendation model;
and step ten, recommending the articles for the group to be recommended by using the group recommendation model after training.
Further, the second step includes:
step two A, training data and a user article relation diagramInputting a user recommendation model, and randomly initializing a vector +.>As user->Is to randomly initialize a vector +.>As an article->Is an initial embedded representation of an item;
step two B, initializing a userEmbedded representation of a convolution layer at layer 0 of the user recommendation model +.>Initializing the article->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>The method comprises the steps of carrying out a first treatment on the surface of the Define user +.>In user recommendation model->The embedding of the layer convolution layer is denoted +.>Define the article->In user recommendation model->The embedding of the layer convolution layer is denoted +.>;
Step two, C, userIn user recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
article and method for manufacturing the sameIn user recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
wherein,convolving the normalized coefficients for the user,>representative and user->Interactive item set->Representative and article->A set of users that interact;
step two D, the user is given the following stepsThe embedded representations of the convolution layers of each layer of the user recommendation model are fused to obtain the user +.>User-embedded representation +.>:
;
Article to be processedEmbedding, representing and fusing all layers of convolution layers of user recommendation model to obtain article +.>Item embedded representation of (c):
;
Wherein,represents the total number of convolutions in the user recommendation model, < >>Representing a polymerization process.
Further, the third step specifically includes:
quantification of users by means of dot product calculationFor articles->Is +.>:
;
Wherein,representing a matrix transpose operation.
Further, in the fourth step, the loss function based on the user personalized recommendationSpecifically, it refers to:
;
wherein,for L2 regularization formula, +.>For user->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>And (2) is->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>Set of->Representative and user->Interactive item set->Representative article->Is->Interaction occurs, ->Representative article->Is->No interaction occurs->Representing a sigmoid function->Representing natural logarithmic sign, < >>Representing user +.>For articles->Is a preferred degree of (a) is a preferred degree of (b).
Further, the fifth step specifically includes:
using aggregation methods to group populationsUser preferences for all group members are embedded and aggregated into a group/>Population initial embedding->:
;
Wherein,representative group->Is a group member collection.
Further, the sixth step specifically includes:
group of peopleAnd (3) with the crowd->Preference similarity between->The method comprises the following steps:
;
wherein,representative group->Is a group member collection->Representative group->Is a group member collection->Representing a union operation, ++>Representing an absolute value operation.
Further, the seventh step specifically includes:
step seven A, group is formedPopulation initial embedding->Article->Is embedded with item characteristics->Group->And (3) with the crowd->Preference similarity between->And group article relationship diagram->Inputting a group recommendation model convolution layer to enable the group +.>Population initial embedding->As a group->Is an initial embedded representation of, item->Is embedded with item characteristics->As an article->Is used for embedding the initial embedded vector of the (a);
step seven B, initializing a populationEmbedded representation of the convolution layer at layer 0 of the population recommendation model +.>Initializing the article->Embedding vector in layer 0 convolution layer of population recommendation model +.>The method comprises the steps of carrying out a first treatment on the surface of the Define group->At group recommendation model->The embedding of the layer convolution layer is denoted +.>Define the article->At group recommendation model->The embedding vector of the layer convolution layer is +.>;
Step seven C, populationIn groupBody recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
article and method for manufacturing the sameAt group recommendation model->Embedding vector of layer convolution layer->The method comprises the following steps:
;
wherein,to control the hyper-parameters of population preference specific gravity of mutually similar groups in convolution propagation, +.>Normalized coefficients for population convolution ++>For->Interactive item set, < > for>For +.>A collection of groups that interact with each other,representative group->Is a set of similarity groups->Representative group->Is->At group recommendation model->An embedded representation of the layer convolution layer;
step seven D, groupEmbedding, representing and fusing all layers of convolution layers in a group recommendation model to obtain a group +.>Group embedded representation->:
;
Article to be processedEmbedding vector fusion of convolution layers of each layer of the group recommendation model to obtain an article +.>Is an article embedded vector of (a):
;
Wherein,representing the total number of layers of the convolution layers in the population recommendation model.
Further, the eighth step specifically includes:
quantification of populations by means of dot product calculationFor articles->Is +.>:
。
Further, step nine is a loss function based on population recommendationThe method comprises the following steps:
;
wherein,for->Interactive item set, < > for>Representative article->And (3) with the crowd->The interaction is made to take place such that,representative article->And (3) with the crowd->No interaction occurs->For crowd->Embedded representation of the convolution layer at layer 0 of the population recommendation model +.>And (2) is->Embedding vector in layer 0 convolution layer of population recommendation model +.>Set of->For crowd->For articlesIs (are) preferred by->For L2 regularization formula, +.>Representing a sigmoid function->Representing natureLogarithmic sign.
One or more technical solutions provided in the embodiments of the present application at least have the following technical effects or advantages:
compared with the prior art, the group recommendation method based on graph convolution comprises group embedding pre-training and heuristic group aggregation strategies based on user history interaction data, group preference similarity calculation based on cosine similarity and group high-order interaction information diffusion based on graph convolution neural network. The building model simulates the high-order interaction and social propagation process of the groups by using the graph convolution neural network, calculates the preference similarity between the groups based on cosine similarity by using common group members between the groups according to the sparsity of the group interaction, enriches the group preference information based on the preference similarity between the groups, and therefore improves the group recommendation accuracy of the model, and simultaneously maintains the recommendation efficiency.
Drawings
Fig. 1 is a schematic diagram of a group movie recommendation application scene provided in the background art of the present invention;
FIG. 2 is a schematic diagram of enriching group preference information based on preference similarity of group common group members provided in the background of the invention;
FIG. 3 is a flowchart of a group recommendation method based on graph convolution according to an embodiment of the present invention;
fig. 4 is a model structure diagram of a group recommendation model according to an embodiment of the present invention.
Detailed Description
The present invention will be described in detail below with reference to the drawings and detailed embodiments, and before the technical solutions of the embodiments of the present invention are described in detail, the terms and terms involved will be explained, and in the present specification, the components with the same names or the same reference numerals represent similar or identical structures, and are only limited for illustrative purposes.
The invention provides a group recommendation method based on graph convolution, which uses a graph convolution neural network to simulate a high-order interaction relation among diffusion groups, uses a heuristic average aggregation strategy to aggregate preference information of group members, uses group members shared among groups to calculate similarity among the groups, and uses group members and article interaction histories to calculate similarity among articles so as to enhance the capability of the high-order interaction diffusion of the groups.
A group recommendation method based on graph convolution is shown in fig. 3, and comprises three parts: embedding a pre-training and heuristic group aggregation strategy based on the group of the user history interaction data; group preference similarity calculation based on cosine similarity; group high-order mutual information diffusion based on graph convolution neural network. Based on the method, the invention constructs a user recommendation model and a group recommendation model, and the structures of the two models are shown in figure 4.
1. Group embedding pre-training and heuristic group aggregation strategy based on user history interaction data
Using user item relationship graphsRepresenting the interaction relationship between the user and the object, user object relationship diagram +.>As undirected graph, user item relationship graph +.>Each node representing a user or item, each edge connecting a user with an item, each edge representing interactions between the user connected to the edge and the item connected to the edge.
Using group article relationship graphsRepresenting interaction relationship between group and article, group article relationship diagram +.>Is undirected graph, group article relation graph +.>Each node of (a) represents a group or article, each edge connects a group with an article,each edge represents interactions between the group to which the edge is attached and the item to which the edge is attached.
The invention adopts the data of the online life recommendation platform as training data, such as a public comment platform and a bean cotyledon platform. User information and interaction information of users and articles are collected from data of the platform, and if a group of users with social connection have coincidence activity data at the same time and the same place, the users are regarded as members of a group, and the coincidence activity data are regarded as interaction data of the group.
In order to utilize the personal preference information in a user recommendation model, the historical interaction data of the user and the article is converted into user embedding and article embedding in a pre-training mode.
Specifically, the present invention uses a graph convolutional neural network to pretrain to obtain user embeddings and item embeddings. Mapping training data to user itemsInputting a user recommendation model, and randomly initializing a vector +.>As user->Is to randomly initialize a vector +.>As an article->Is to initialize the user +.>Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>InitializingArticle->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>Define user +.>In user recommendation model->The embedding of the layer convolution layer is denoted +.>Define the article->In user recommendation model->The embedding of the layer convolution layer is denoted +.>User +.>In user recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
article and method for manufacturing the sameIn user recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
wherein,for the user convolution normalization coefficient, the user convolution normalization coefficient ensures that the embedded scale does not increase with the increase of the convolution layer number; />Representative and user->Interactive item set->Representative and article->A collection of users who interact.
Will userThe embedded representations of the convolution layers of each layer of the user recommendation model are fused to obtain the user +.>Is a user-embedded representation of (a):
;
Article to be processedEmbedding table of convolution layers of user recommendation modelShowing fusion to get article->Item embedded representation of (c):
;
Wherein,represents the total number of convolutions in the user recommendation model, < >>Representative polymerization processes, the present invention employs an average polymerization process.
Obtaining the userUser-embedded representation +.>And (2) is->Item embedded representation->After that, the user is quantized by means of dot product calculation>For articles->Is +.>:
;
Wherein,representing a matrix transpose operation.
Loss function based on user-personalized recommendations using Bayesian Personalized Ranking (BPR) loss definitionPre-training a user recommendation model to obtain user +.>Is embedded with user preferences->And articles->Is embedded with item characteristics->Loss function based on user personalized recommendation +.>The method comprises the following steps:
;
wherein,for L2 regularization formula, +.>For user->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>And (2) is->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>Set of->Representative article->Is->Interaction occurs, ->Representative article->Is->No interaction occurs->Representing a sigmoid function->Representing natural logarithmic sign, < >>Representing user +.>For articles->Is a preferred degree of (a) is a preferred degree of (b).
The decision made by the population is largely determined by the preferences of the population members, and in order to take advantage of the personalized preferences of the population members in the population recommendation, the user preferences of all population members are embedded and aggregated into a population using an aggregation methodPopulation initial embedding->:
;
Wherein,representative group->Is a group member collection.
Polymerization processThere are various options such as maximum satisfaction, minimum pain, average or attentive neural networks, etc., and the invention also uses an average aggregation method in view of the simplicity of the user recommendation model and the efficiency of training and prediction.
2. Group preference similarity calculation based on cosine similarity
There may be common population members, definitions and populations between different populationsPopulation ∈with members of the common population>For crowd->To some extent, these common population members reveal a common interest between the populations that are similar to each other. The preference information of the similar groups is added into the embedded representation of the group, so that the group preference information capturing capability of the group recommendation model can be effectively improved. Calculating population ∈by common population members>And (3) with the crowd->Cosine similarity between them to obtain a populationAnd (3) with the crowd->Preference similarity between->:
;
Wherein,representative group->Is a group member collection->Representing a union operation, ++>Representing an absolute value operation.
3. Group high-order interaction information diffusion based on graph convolution neural network
The historical interaction of the group and the article comprises the preference information of the group, and the preference information of the group hidden in the interaction of the group and the article can be obtained in a picture convolution mode, however, the historical interaction of the group and the article is far more sparse than the historical interaction of the user and the article, which means that the relationship graph of the group article is simply usedThe information promotion effect brought by group recommendation is far better than that of using the user object relation graph +.>The information promotion effect on the user recommendation is small, and the difficulty of group recommendation is also the same.
Therefore, the invention combines the preference similarity among the groups with the graph convolution, so that the information of the groups is simultaneously diffused into the embedded representation of the similar groups in the diffusion process of the graph convolution, thereby effectively filling the relationship graph of the group objectsThe adverse effect of sparseness.
Will be the populationPopulation initial embedding->Article->Is embedded with item characteristics->Group->And (3) with the crowd->Preference similarity between->And group article relationship diagram->And inputting a group recommendation model. Will crowd->Initial embedding of population of (a)As a group->Is embedded initially in (a)Indicating, article->Is embedded with item characteristics->As an article->Is used to determine the initial embedded vector of (a). Initializing group->Embedded representation of the convolution layer at layer 0 of the population recommendation model +.>Initializing the article->Embedding vector in layer 0 convolution layer of population recommendation model +.>Define the group->At group recommendation model->The embedding of the layer convolution layer is denoted +.>Define the article->At group recommendation model->The embedding vector of the layer convolution layer is +.>Group->At group recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
article and method for manufacturing the sameAt group recommendation model->Embedding vector of layer convolution layer->The method comprises the following steps:
;
wherein,to control the hyper-parameters of population preference specific gravity of mutually similar groups in convolution propagation, +.>For the group convolution normalization coefficient, ensure that the embedded scale does not increase with the increase of the convolution layer number,/->For->Interactive item set, < > for>For +.>Interactive population collection, +.>Representative group->Is a set of similarity groups->Representative group->Is->At group recommendation model->An embedded representation of a layer convolution layer.
Will be the populationEmbedding, representing and fusing all layers of convolution layers in a group recommendation model to obtain a group +.>Group embedded representation->:
;
Article to be processedEmbedding vector fusion of convolution layers of each layer of the group recommendation model to obtain an article +.>Is an article embedded vector of (a):
;
Wherein,representing the total number of layers of the convolution layers in the population recommendation model. The invention takes->The total layer number of the convolution layers in the user recommendation model is equal to the total layer number of the convolution layers in the group recommendation model.
Similarly, the polymerization process in the above formulaAn average polymerization method is used.
Obtaining a populationGroup embedded representation->And (2) is->Item-embedding vector +.>Afterwards, the population is quantified by dot product calculation>For articles->Is +.>:
。
Loss function based on group recommendations using Bayesian Personalized Ranking (BPR) loss definitionTraining the group recommendation model to obtain a trained group recommendation model, wherein the loss function is based on group recommendation +.>The method comprises the following steps:
;
wherein,representative article->And (3) with the crowd->Interaction occurs, ->Representative article->And (3) with the crowd->No interaction occurs->For crowd->Embedded representation of the convolution layer at layer 0 of the population recommendation model +.>And (2) is->Group recommendationEmbedding vector of model layer 0 convolution layer +.>Set of->For crowd->For articles->Is a preferred degree of (a) is a preferred degree of (b).
Taking a plurality of articles which are not interacted with the group to be recommended as a group of articles to be recommended, predicting the preference degree of all articles in the group to be recommended by using a group recommendation model which completes training, sequencing the obtained preference degree, and taking the sequenced result as a groupAnd recommending.
The above embodiments are merely illustrative of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention, and various modifications and improvements made by those skilled in the art to the technical solution of the present invention should fall within the protection scope defined by the claims of the present invention without departing from the design spirit of the present invention.
Claims (7)
1. The group recommending method based on graph convolution is characterized by comprising the following steps of:
step one, preparing user and article interaction data and group and article interaction data as training data, wherein the training data comprises usersArticle->Group->And (3) with the crowd->The method comprises the steps of carrying out a first treatment on the surface of the Use of user item relationship diagram->Representing the interaction relationship between the user and the article; use of group article relationship diagram->Representing interaction relationship between the group and the article;
step two, training data and a user article relation diagramInputting a user recommendation model convolution layer to obtain a user +.>User-embedded representation +.>And articles->Item embedded representation->The method comprises the steps of carrying out a first treatment on the surface of the The second step comprises:
step two A, training data and a user article relation diagramInputting a user recommendation model, and randomly initializing a vector +.>As user->Is to randomly initialize a vector +.>As an article->Is an initial embedded representation of an item;
step two B, initializing a userEmbedded representation of a convolution layer at layer 0 of the user recommendation model +.>Initializing the article->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>The method comprises the steps of carrying out a first treatment on the surface of the Define user +.>In user recommendation model->The embedding of the layer convolution layer is denoted +.>Define the article->In user recommendation model->The embedding of the layer convolution layer is denoted +.>;
Step two, C, userIn user recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
article and method for manufacturing the sameIn user recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
wherein,convolving the normalized coefficients for the user,>representative and user->Interactive item set->Representative and articleA set of users that interact;
step two D, the user is given the following stepsThe embedded representations of the convolution layers of each layer of the user recommendation model are fused to obtain the user +.>User-embedded representation +.>:
;
Article to be processedEmbedding, representing and fusing all layers of convolution layers of user recommendation model to obtain article +.>Item embedded representation->:
;
Wherein,represents the total number of convolutions in the user recommendation model, < >>Represents a polymerization process;
step three, according to the userUser-embedded representation +.>And articles->Item embedded representation->Quantification of user +.>For articles->Is +.>;
Step four, defining a loss function based on user personalized recommendationPre-training a user recommendation model to obtain user +.>Is embedded with user preferences->And articles->Is embedded with item characteristics->;
Step five, using a polymerization methodWill be the populationUser preferences of all group members of (a) are embedded and aggregated into +.>Population initial embedding->;
Step six, calculating the population through the common population membersAnd (3) with the crowd->Cosine similarity between them gives the population +.>And (3) with the crowd->Preference similarity between->;
Step seven, the group is formedPopulation initial embedding->Article->Is embedded with item characteristics->Group->And (3) with the crowd->Preference similarity between->And group article relationship diagram->Inputting a group recommendation model convolution layer to obtain a group +.>Group embedded representation->And articles->Item-embedding vector +.>The method comprises the steps of carrying out a first treatment on the surface of the The seventh step specifically comprises:
step seven A, group is formedPopulation initial embedding->Article->Is embedded with item characteristics->Group->And (3) with the crowd->Preference similarity between->And group article relationship diagram->Inputting a group recommendation model convolution layer to enable the group +.>Initial embedding of population of (a)As a group->Is an initial embedded representation of, item->Is embedded with item characteristics->As an article->Is used for embedding the initial embedded vector of the (a);
step seven B, initializing a populationEmbedded representation of the convolution layer at layer 0 of the population recommendation model +.>Initializing the article->Embedding vector in layer 0 convolution layer of population recommendation model +.>The method comprises the steps of carrying out a first treatment on the surface of the Define group->At group recommendation model->The embedding of the layer convolution layer is denoted +.>Define the article->At group recommendation model->The embedding vector of the layer convolution layer is +.>;
Step seven C, populationAt group recommendation model->Embedded representation of layer convolution layer->The method comprises the following steps:
;
article and method for manufacturing the sameAt group recommendation model->Embedding vectors for layer convolution layers/>The method comprises the following steps:
;
wherein,to control the hyper-parameters of population preference specific gravity of mutually similar groups in convolution propagation, +.>Normalized coefficients for population convolution ++>For->Interactive item set, < > for>For +.>Interactive population collection, +.>Representative populationIs a set of similarity groups->Representative group->Is->At group recommendation model->An embedded representation of the layer convolution layer;
step seven D, groupEmbedding, representing and fusing all layers of convolution layers in a group recommendation model to obtain a group +.>Group embedded representation->:
;
Article to be processedEmbedding vector fusion of convolution layers of each layer of the group recommendation model to obtain an article +.>Item-embedding vector +.>:
;
Wherein,representing the total layer number of the convolution layers in the group recommendation model;
step eight, according to the groupGroup embedded representation->And articles->Item-embedding vector +.>Quantification of the population by dot product calculation>For articles->Is +.>;
Step nine, defining a loss function based on group recommendationTraining the group recommendation model to obtain a trained group recommendation model;
and step ten, recommending the articles for the group to be recommended by using the group recommendation model after training.
2. The population recommendation method based on graph convolution as claimed in claim 1, wherein the third step specifically comprises:
quantification of users by means of dot product calculationFor articles->Is +.>:
;
Wherein,representing a matrix transpose operation.
3. The graph convolution-based group recommendation method according to claim 1, wherein the loss function based on user personalized recommendation in the fourth stepSpecifically, it refers to:
;
wherein,for L2 regularization formula, +.>For user->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>And (2) is->Embedded representation of a convolution layer at layer 0 of the user recommendation model +.>Set of->Representative and user->Interactive item set->Representative article->Is->Interaction occurs, ->Representative article->Is->No interaction occurs->Representing a sigmoid function->Representing natural logarithmic sign, < >>Representing user +.>For articles->Is a preferred degree of (a) is a preferred degree of (b).
4. The population recommendation method based on graph convolution as claimed in claim 1, wherein the fifth step specifically comprises:
using aggregation methods to group populationsUser preferences of all population members are embedded and aggregated into a population +.>Initial embedding of population of (a):
;
Wherein,representative group->Is a group member collection.
5. The population recommendation method based on graph convolution as claimed in claim 1, wherein the sixth step specifically comprises:
group of peopleAnd (3) with the crowd->Preference similarity between->The method comprises the following steps:
;
wherein,representative group->Is a group member collection->Representative group->Is a group member collection->Representing a union operation, ++>Representing an absolute value operation.
6. The population recommendation method based on graph convolution as claimed in claim 1, wherein the step eight specifically comprises:
quantification of populations by means of dot product calculationFor articles->Is +.>:
。
7. The method of claim 1, wherein in step nine the group-based recommendation method is performed byLoss function for volume recommendationThe method comprises the following steps:
;
wherein,for->Interactive item set, < > for>Representative article->And (3) with the crowd->Interaction occurs, ->Representative article->And (3) with the crowd->No interaction occurs->For crowd->Embedded representation of the convolution layer at layer 0 of the population recommendation model +.>And articlesEmbedding vector in layer 0 convolution layer of population recommendation model +.>Set of->For crowd->For articles->Is provided with a degree of preference for (a),for L2 regularization formula, +.>Representing a sigmoid function->Representing natural logarithmic symbols.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311374128.XA CN117112914B (en) | 2023-10-23 | 2023-10-23 | Group recommendation method based on graph convolution |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311374128.XA CN117112914B (en) | 2023-10-23 | 2023-10-23 | Group recommendation method based on graph convolution |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117112914A CN117112914A (en) | 2023-11-24 |
CN117112914B true CN117112914B (en) | 2024-02-09 |
Family
ID=88796964
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202311374128.XA Active CN117112914B (en) | 2023-10-23 | 2023-10-23 | Group recommendation method based on graph convolution |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN117112914B (en) |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160954A (en) * | 2019-12-16 | 2020-05-15 | 南京理工大学 | Recommendation method facing group object based on graph convolution network model |
CN113407864A (en) * | 2021-06-21 | 2021-09-17 | 南京邮电大学 | Group recommendation method based on mixed attention network |
CN113918832A (en) * | 2021-10-22 | 2022-01-11 | 重庆理工大学 | Graph convolution collaborative filtering recommendation system based on social relationship |
CN114741572A (en) * | 2022-04-08 | 2022-07-12 | 山东省人工智能研究院 | Group recommendation method based on graph convolution neural network group discovery |
CN116204729A (en) * | 2022-12-05 | 2023-06-02 | 重庆邮电大学 | Cross-domain group intelligent recommendation method based on hypergraph neural network |
CN116797312A (en) * | 2023-05-10 | 2023-09-22 | 北京林业大学 | Group recommendation method integrating lightweight graph convolution network and attention mechanism |
CN116894122A (en) * | 2023-07-06 | 2023-10-17 | 黑龙江大学 | Cross-view contrast learning group recommendation method based on hypergraph convolutional network |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11443346B2 (en) * | 2019-10-14 | 2022-09-13 | Visa International Service Association | Group item recommendations for ephemeral groups based on mutual information maximization |
KR20230044885A (en) * | 2021-09-27 | 2023-04-04 | 삼성전자주식회사 | SYSTEM AND METHOD FOR PROVIDING recommendation contents |
EP4202725A1 (en) * | 2021-12-22 | 2023-06-28 | Naver Corporation | Joint personalized search and recommendation with hypergraph convolutional networks |
-
2023
- 2023-10-23 CN CN202311374128.XA patent/CN117112914B/en active Active
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111160954A (en) * | 2019-12-16 | 2020-05-15 | 南京理工大学 | Recommendation method facing group object based on graph convolution network model |
CN113407864A (en) * | 2021-06-21 | 2021-09-17 | 南京邮电大学 | Group recommendation method based on mixed attention network |
CN113918832A (en) * | 2021-10-22 | 2022-01-11 | 重庆理工大学 | Graph convolution collaborative filtering recommendation system based on social relationship |
CN114741572A (en) * | 2022-04-08 | 2022-07-12 | 山东省人工智能研究院 | Group recommendation method based on graph convolution neural network group discovery |
CN116204729A (en) * | 2022-12-05 | 2023-06-02 | 重庆邮电大学 | Cross-domain group intelligent recommendation method based on hypergraph neural network |
CN116797312A (en) * | 2023-05-10 | 2023-09-22 | 北京林业大学 | Group recommendation method integrating lightweight graph convolution network and attention mechanism |
CN116894122A (en) * | 2023-07-06 | 2023-10-17 | 黑龙江大学 | Cross-view contrast learning group recommendation method based on hypergraph convolutional network |
Non-Patent Citations (3)
Title |
---|
基于注意力机制的神经网络贝叶斯群组推荐算法;李诗文;潘善亮;;计算机应用与软件(第05期);全文 * |
基于注意力机制的神经网络贝叶斯群组推荐算法;杨远奇;;数字技术与应用(第08期);全文 * |
群组推荐系统研究与分析;王金;张朝恒;;计算机技术与发展(第05期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN117112914A (en) | 2023-11-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021203819A1 (en) | Content recommendation method and apparatus, electronic device, and storage medium | |
CN111428147B (en) | Social recommendation method of heterogeneous graph volume network combining social and interest information | |
CN111931062B (en) | Training method and related device of information recommendation model | |
US20220171760A1 (en) | Data processing method and apparatus, computer-readable storage medium, and electronic device | |
CN109871485A (en) | A kind of personalized recommendation method and device | |
CN109034960B (en) | Multi-attribute inference method based on user node embedding | |
CN111428127B (en) | Personalized event recommendation method and system integrating theme matching and bidirectional preference | |
CN109101629A (en) | A kind of network representation method based on depth network structure and nodal community | |
CN114202061A (en) | Article recommendation method, electronic device and medium based on generation of confrontation network model and deep reinforcement learning | |
CN107103499A (en) | Method is recommended in a kind of cross-site cold start-up based on deep learning | |
CN109325875A (en) | Implicit group based on the hidden feature of online social user finds method | |
CN113343119A (en) | Group recommendation method based on hierarchical attention mechanism | |
Chen et al. | CoNet: Co-occurrence neural networks for recommendation | |
CN113379494A (en) | Commodity recommendation method and device based on heterogeneous social relationship and electronic equipment | |
CN112559878A (en) | Sequence recommendation system and recommendation method based on graph neural network | |
CN116204729A (en) | Cross-domain group intelligent recommendation method based on hypergraph neural network | |
Murrugarra-Llerena et al. | Image retrieval with mixed initiative and multimodal feedback | |
CN117112914B (en) | Group recommendation method based on graph convolution | |
KR20210079264A (en) | Methods and systems for defining emotional machines | |
CN117251586A (en) | Multimedia resource recommendation method, device and storage medium | |
Yang et al. | A collaborative filtering recommendation based on user profile and user behavior in online social networks | |
Qader et al. | Dual-stage social friend recommendation system based on user interests | |
Nosshi et al. | Hybrid recommender system via personalized users’ context | |
CN114547480A (en) | Deep learning recommendation method and system based on multi-platform fusion | |
Wan et al. | A learner’s role-based multi dimensional collaborative recommendation (LRMDCR) for group learning support |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |