CN112686736B - System recommendation method - Google Patents

System recommendation method Download PDF

Info

Publication number
CN112686736B
CN112686736B CN202110027315.5A CN202110027315A CN112686736B CN 112686736 B CN112686736 B CN 112686736B CN 202110027315 A CN202110027315 A CN 202110027315A CN 112686736 B CN112686736 B CN 112686736B
Authority
CN
China
Prior art keywords
user
network
representing
commodity
vector
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
Application number
CN202110027315.5A
Other languages
Chinese (zh)
Other versions
CN112686736A (en
Inventor
陈卓
李涵
杜军威
魏锐
姜伟豪
葛艳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Qingdao University of Science and Technology
Original Assignee
Qingdao University of Science and Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Qingdao University of Science and Technology filed Critical Qingdao University of Science and Technology
Priority to CN202110027315.5A priority Critical patent/CN112686736B/en
Publication of CN112686736A publication Critical patent/CN112686736A/en
Application granted granted Critical
Publication of CN112686736B publication Critical patent/CN112686736B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a system recommendation method, which comprises the following specific processing steps: 101) a data relation association step, 102) a user relation data fusion step, 103) a commodity data fusion step, and 104) a recommendation score prediction step; the invention provides a system recommendation method for performing representation learning on nodes through a heterogeneous network, enhancing node feature representation and improving sparse data processing capacity.

Description

System recommendation method
Technical Field
The invention relates to the technical field of recommendation, in particular to a system recommendation method.
Background
With the development of the internet and big data technology, an information explosion age has come, various information is presented to people, and as a technology for helping people to select useful information, a recommendation system has been presented in various fields of medical treatment, business, education, etc., and shows a strong ability.
With the development of the times, network data become more and more complex, data in a network is not composed of a single type of elements, a large amount of different types of data form various network graphs, and a traditional recommendation algorithm can only process a regular topological graph composed of a single type of data, so that a Heterogeneous Information Network (HIN) is produced. Heterogeneous information networks have been applied to the recommendation field as a powerful modeling method. A heterogeneous information network is a special type of network that contains a large number of different types of nodes connected together by different types of edges that represent different relationships between network nodes. The heterogeneous information network is integrated into a recommendation system, and the development of the field is accelerated to a great extent.
In recent years, the graph neural network is also widely applied to a recommendation system, and has the function of fusing neighbor information of nodes to show more characteristic information of the nodes, but the graph neural network has defects of the graph neural network, and the graph Rec model is only a single first-order neighbor fused with the nodes, and cannot extract deeper relationships of the nodes. If the data set is sparse, the number of neighbor nodes of the node is relatively small, the feature information of the node cannot be fully fused, and therefore prediction inaccuracy is increased.
Disclosure of Invention
The invention solves the technical problems and provides a system recommendation method for performing representation learning on nodes through a heterogeneous network, enhancing node feature representation and improving sparse data processing capacity.
The technical scheme of the invention is as follows:
a system recommendation method comprises the following specific processing steps:
101) data relation association step: by U ═ U1,u2,...,unDenotes a user data set, V ═ V1,v2,...,vmThe representation of the data set of the goods,
Figure BDA0002890781840000021
representing meta-paths between data, a representing a user or a commodity, R representing a social relationship or a buy and sell relationship; the meta path is generated as follows:
Figure BDA0002890781840000022
wherein n istIs the current node, nt+1In order to be the next node, the node is,
Figure BDA0002890781840000023
among the neighbors representing v are those belonging to At+1ρ represents a meta path rule, and the length of the meta path is set to be L; phi (x) represents the type of the next node of the path, belonging to the node type in the network; v and x represent the current node and the next node, respectively, belonging to all nodes in the network;
meta-path of user and commodity interactive relationThe generated vector of the user and the commodity is expressed as
Figure BDA0002890781840000024
And
Figure BDA0002890781840000025
the user vector generated by the meta path of the user social relationship is represented as
Figure BDA0002890781840000026
Score vector representation er
102) And (3) user relationship data fusion: the user data set comprises u _ v of a user item interactive network and u _ u of a user social network; the u _ v of the user item interactive network is combined with the attribute network to generate corresponding weight to obtain the implicit vector representation of the user in the user interactive network, and the implicit vector representation is specifically represented by the following formula:
Figure BDA0002890781840000027
wherein,
Figure BDA0002890781840000028
representing the implicit vector of a user i in an interactive network, sigma being an activation function, w and b being parameters of a neural network, C representing all scores of the user on the commodities, and NiDenotes the neighbors of node i, softmax denotes the softmax function, xjrVector fusion, g, representing item j and corresponding score rτIs a multi-layer fully-connected neural network,
Figure BDA0002890781840000031
representing a concatenation of vectors, w1、w2、b1、b2Representing parameters of two fully connected layers of an attention network;
generating corresponding weight by combining u _ u of the user social network with the attention network to obtain implicit vector representation of the user in the social network, wherein the implicit vector representation is specifically represented by the following formula:
Figure BDA0002890781840000032
wherein,
Figure BDA0002890781840000033
representing an implicit vector of a user i in the social network, and i' representing a neighbor user;
the user vector formula that ultimately includes the user project interaction network and the user social network is as follows:
Figure BDA0002890781840000034
hirepresenting the end-user vector,/n-1Indicates the number of network layers, wn、bnParameters representing w and b multilayer neural networks;
103) and (3) commodity data fusion: the deep relationship of the commodity in the network is collected in a meta-path mode, and the feature vector of the commodity is represented by fusing the features of the neighbor users through GNN
Figure BDA0002890781840000035
The specific formula is as follows:
Figure BDA0002890781840000036
wherein,
Figure BDA0002890781840000037
representing an implicit vector of the commodity j in the interactive network;
104) and (3) recommendation score prediction step: predicting the scores of the user on the commodities according to the user vectors and the commodity vectors obtained in the steps 101) to 103), splicing the user vectors and the commodity vectors through a full-connection layer network to serve as the input of the network, and using the predicted scores as the output of the network, namely using a multilayer full-connection neural network to obtain the final predicted scores, wherein the specific formula is as follows:
Figure BDA0002890781840000041
wherein,
Figure BDA0002890781840000042
is the prediction score.
Further, the specific process of implicit vector representation of the user in the user interaction network in step 102) is as follows:
Figure BDA0002890781840000043
where aggregate represents the fusion function, xjrThe vectors representing the commodities j and the corresponding scores r are fused, and C represents all scores which are made by the user on the commodities; x is the number ofjrThe formula of (1) is as follows:
Figure BDA0002890781840000044
gτis a multi-layer fully-connected neural network,
Figure BDA0002890781840000045
representing the concatenation of the vectors;
the Aggregation fusion function adopts a fully-connected neural network, and the concrete formula is as follows:
Figure BDA0002890781840000046
σ is the activation function, w and b are neural network parameters, αjiFor the weight of a commodity j to a user i, the importance degree of each commodity to the user is different, the weight is trained through an Attention network, and the specific formula is as follows:
Figure BDA0002890781840000047
further, the implicit vector of the user in the social network of the user in step 102) is represented by the neighboring vector fusion of the user, and specifically, the training weight process in combination with the Attention network is as follows:
Figure BDA0002890781840000048
Figure BDA0002890781840000051
wherein beta isi′iThe weight index of the neighbor i' to the user i is represented by the following specific formula:
Figure BDA0002890781840000052
further, the method also comprises 105) a parameter adjusting step: adjusting parameters of the recommendation model established in the steps 101) to 104), and specifically, defining an objective function to optimize so as to reversely update the parameters in the recommendation model; the Adam method is selected for adjustment and optimization, wherein the adjusted loss formula is as follows:
Figure BDA0002890781840000053
wherein,
Figure BDA0002890781840000054
predictive score, y, for user i for item jijFor a true score, n + m is the total number of users and goods.
Compared with the prior art, the invention has the advantages that:
the invention has the advantage of processing data with increased sparsity. The scheme model introduces a method for generating vectors by heterogeneous network element paths on the basis of a graph neural network, and learns the feature representation of the enhanced node through the heterogeneous network representation. Tests are carried out on the two public data sets, the results show that the model is superior to other models, and the model is proved to be more competitive in solving the sparse problem by changing the sparsity of the data sets. At present, a large number of data sets on a network contain a plurality of node attribute elements, and the attributes are fully utilized to effectively help a model and accurate prediction.
Drawings
FIG. 1 is a frame diagram of the present invention;
FIG. 2 is a user merchandise interaction data diagram of the present invention;
FIG. 3 is a diagram of a user's social network of the present invention;
FIG. 4 is a first set of learning rate experimental statistics of the present invention;
FIG. 5 is a second set of learning rate experimental statistics of the present invention;
FIG. 6 is a comparison graph of the cia dataset embedding dimension experiment index RMSE of the present invention;
FIG. 7 is a comparison graph of the ciao data set embedding dimension experiment index MAE of the present invention.
Detailed Description
The present invention is further described in the following detailed description in conjunction with the drawings, and portions not described or illustrated in detail herein can be implemented using conventional technology.
As shown in fig. 1 to 7, a system recommendation method includes the following specific processing steps:
101) data relation association step: from U ═ U1,u2,...,unDenotes a user data set, and n denotes the number of users. V ═ V1,v2,...,vmDenotes a commodity data set, and m denotes the number of commodities. P represents a meta-path between data, the specific meta-path being
Figure BDA0002890781840000061
Wherein, A represents the entity in the network diagram, namely the user or the commodity, and R represents the interactive relationship between the entities, namely the social relationship or the buying and selling relationship;T∈Rn*mfor the user commodity interaction matrix, rijRepresents the grade given by the user i to the commodity j, and r is given if the user does not mark the gradeijThe score may be regarded as 0, and the higher the score is, the more the user likes the product. One user can score multiple items, as can one item scored by multiple users. G is belonged to Rn*nA social matrix for the user. And delta-R represents a network structure diagram, and is divided into a user commodity interaction diagram u _ v and a user social relationship diagram u _ u. The meta path is generated as follows:
Figure BDA0002890781840000062
wherein n istIs the current node, nt+1In order to be the next node, the node is,
Figure BDA0002890781840000063
among the neighbors representing v are those belonging to At+1ρ represents a meta path rule, and the length of the meta path is set to L. Phi (x) represents the type of the next node of the path, belonging to the node type in the network; v and x represent the current node and the next node, respectively, belonging to all nodes in the network.
Then, a large number of path instances are generated according to the generation mode, and the vectors of the nodes generated by the classical model word2vec are substituted for expression, wherein the specific formula is as follows:
Figure BDA0002890781840000071
wherein context and NEG respectively represent positive and negative sample sets of the sample w, the window size is set to be related to the size of the positive sample, and the window size minus one is the number of the positive samples. The positive samples of w are selected from the path examples, specifically refer to a certain number of nodes, and are related to the path examples and the window size, and the selected positive samples are all contained in the path examples; the negative sample set is a node selected in addition to the nodes in the positive sample, and may or may not be in the path. The window size determines the number of nodes associated with the node, and the larger the window, the greater the number of associated nodes selected. For example, one example of a path is u1 → v1 → u2 → v2 → u3 → v3 → u4 → v4, and when the window size is 3, the positive samples are u3 and u4 for the v3 node, and the number is 2.
The vector representation of the user and the commodity generated by the meta path of the interaction relation of the user and the commodity is
Figure BDA0002890781840000072
And
Figure BDA0002890781840000073
the user vector generated by the meta path of the user social relationship is represented as
Figure BDA0002890781840000074
Score vector representation er
Specifically, as shown in FIG. 2, one of the meta-paths is u1→v1→u3→v2→u4The path length L is 5. As shown in FIG. 3, one of the meta-paths is u2→u4→u3→u1→u6The path length L is 5. Different meta-paths are generated aiming at different network graphs, and are used as sentences, word2vec is used for generating a feature vector of each node, so that the multi-order neighbor relation of each node is captured, and the defect that the graph neural network cannot be fused with the high-order neighbor relation is overcome.
102) And (3) user relationship data fusion: the user data set comprises u _ v of the user item interaction network and u _ u of the user social network. And generating corresponding weight by combining the u _ v of the user item interactive network with the attribute network to obtain the implicit vector representation of the user in the user interactive network. Namely, the hidden vector of the user for a certain commodity is represented by the neighbor commodity and the score of the user, then corresponding weight is generated through an attention network, and finally all the hidden vectors are fused to obtain the hidden vector representation of the user in the interactive network.
The following formula is shown:
Figure BDA0002890781840000081
wherein,
Figure BDA0002890781840000082
representing the implicit vector of the user i in the interactive network, sigma being an activation function, w and b being parameters of the neural network, C representing all the scores made by the user on the commodity, and NiDenotes the neighbor of node i, softmax denotes the softmax function, xjrVector fusion, g, representing item j and corresponding score rτIs a multi-layer fully-connected neural network,
Figure BDA0002890781840000083
splicing of the representation vectors, w1、w2、b1、b2And a parameter representing two fully connected layers of the attention network, namely the attention network comprises two fully linked layers.
The specific process of implicit vector representation of the user in the user interaction network is as follows:
Figure BDA0002890781840000084
where aggregate represents the fusion function, xjrThe vectors representing the commodities j and the corresponding scores r are fused, and C represents all scores which are made by the user on the commodities; x is the number ofjrThe formula of (1) is as follows:
Figure BDA0002890781840000085
gτis a multi-layer fully-connected neural network,
Figure BDA0002890781840000086
representing the concatenation of the vectors;
the Aggregation fusion function adopts a fully-connected neural network, and the concrete formula is as follows:
Figure BDA0002890781840000087
σ is the activation function, w and b are neural network parameters, αjiFor the weight of a commodity j to a user i, the importance degree of each commodity to the user is different, the weight is trained through an Attention network, two layers of neural networks are used, and the specific formula is as follows:
Figure BDA0002890781840000088
in combination with the above, the user vector is used in the weight calculation, and the final result can be expressed as formula (2):
Figure BDA0002890781840000091
generating corresponding weight by combining u _ u of the social network of the user with the attention network to obtain implicit vector representation of the user in the social network, wherein the implicit vector representation is specifically represented by the following formula:
Figure BDA0002890781840000092
wherein,
Figure BDA0002890781840000093
representing an implicit vector of a user i in the social network, and i' representing a neighbor user;
the implicit vector of the user in the social network of the user is represented, the implicit vector of the user in the social network is represented by the neighbor vector fusion of the user, and the training weight process specifically combined with the Attention network is the following formula:
Figure BDA0002890781840000094
Figure BDA0002890781840000095
wherein beta isi′iThe weight index of the neighbor i' to the user i is represented by the following specific formula:
Figure BDA0002890781840000096
when the weighting index is obtained, the vector representation of the user and the neighbors thereof in the social network is used, so that the final formula is obtained as formula (3):
Figure BDA0002890781840000097
and finally, the user aggregates the interaction relation between the user and the commodity in u _ v, and the user aggregates the social relation between the user and the user in u _ u, and the user vectors of the user project interaction network and the user social network are fused to generate the final user vector. The specific formula is as follows:
Figure BDA0002890781840000098
Figure BDA0002890781840000101
hirepresenting the end user vector, ln-1Indicates the number of network layers, wn、bnParameters representing the w and b neural networks;
103) and (3) commodity data fusion: the deep relationship of the commodity in the network is collected in a meta-path mode, and the feature vector of the commodity is represented by fusing the features of the neighbor users through GNN
Figure BDA0002890781840000102
The concrete formula is as follows:
Figure BDA0002890781840000103
wherein,
Figure BDA0002890781840000104
representing an implicit vector of the commodity j in the interactive network; the other parameters are as defined above.
104) A recommendation score predicting step: the user vector and the commodity vector obtained in steps 101) to 103) are used for predicting the score of the user on the commodity, the splicing of the user vector and the commodity vector is used as the input of the network through a full-connection layer network, the prediction score is used as the output of the network, and the final prediction score is obtained by using a multi-layer (generally, three layers are used here) full-connection neural network. The specific formula is as follows:
Figure BDA0002890781840000105
wherein,
Figure BDA0002890781840000106
is a prediction score. The specific process can be described as first splicing together the final vector representations of the user's goods, then placing them into a multi-layer fully-connected neural network, each layer having an activation function, and finally outputting the predicted value to be obtained.
And 105) parameter adjusting step: adjusting parameters of the recommendation model established in the steps 101) to 104), and specifically, defining an objective function to optimize so as to reversely update the parameters in the recommendation model; the Adam method is selected for adjustment and optimization, wherein the adjusted loss formula is as follows:
Figure BDA0002890781840000111
wherein,
Figure BDA0002890781840000112
predictive score, y, for user i for item jijFor a true score, n + m is the total number of users and goods. The Adam method has the advantages that after offset correction, the learning rate of each iteration has a certain range, so that the parameters are relatively stable. The method provided by the scheme is adopted as a whole, and the loss formula is adopted only for correcting the parameters in the loss adjusting part.
The validity of the model was verified using data on both the Ciao and Epinions websites, crawled from both websites by Asian State university scholars when doing social networking studies. They contained 283319 and 764352 ratings, respectively. And the accuracy of the recommended model is verified by adopting two indexes of RMSE and MAE, and the smaller the value of the two indexes is, the more superior the model is. Meanwhile, other algorithms are compared to show the accuracy of the model of the scheme. Other algorithms for comparison are as follows:
PMF: the main idea of the model is that the user's preference for movies can be determined by a linear combination of a few factors.
SoRec: the main idea is to tie the user's social network relationships to the scoring matrix.
SocialMF: the main idea is to introduce trust propagation in matrix factorization, users representing users that are close to their trust.
NeuMF: the main idea of the model is to combine the traditional matrix decomposition and the multilayer perceptron, and can extract low-dimensional and high-dimensional features at the same time.
DeepsoR: the main idea is that the user representation learned from social relationships is integrated into the probability matrix decomposition.
GraphRec: the main idea is to merge the vectors of the user commodities in different networks for prediction.
Figure BDA0002890781840000113
Figure BDA0002890781840000121
TABLE 1
The sparsity formula for the model is:
Figure BDA0002890781840000122
in the used ciao data set, the sparsity of the data set used by the original model is 98.96%, the number of times of participation and scoring of some users in the data set is large, and the number of times of participation and scoring of some users is small. In order for each user to participate in the interaction, the cold start problem is prevented from occurring. And selecting the users with high scoring participation times to delete the scoring number. Specifically, users with more scoring times of more than or equal to 60 are determined, the problem of cold start of commodities is considered in the deleting process, each commodity is guaranteed to participate in interaction, and finally deleting operation is carried out.
By reducing the user scoring times, the connection relation between the user nodes and the commodity nodes in the heterogeneous network graph is reduced, and the corresponding interaction relation is reduced, so that the field aggregation of the graph neural network cannot be integrated with more features, and finally the recommendation difficulty is increased. In such a sparse case, the effect of the model is detected. Two sets of sparsity variation comparison experiments were performed, with the following results:
Figure BDA0002890781840000123
Figure BDA0002890781840000131
TABLE 2
By reducing the commodities of users with a large number of interactive commodities, the number of scores is reduced, and further the sparsity is increased. 166 scores are deleted to increase the sparsity to 98.98%, 541 scores are deleted to increase the sparsity to 99%, and the sparse data sets are known to be larger through the experimental data sets, so that the model effect of the scheme is improved to be larger compared with that of a contrast model.
Ablation experiments were performed by adding two net representation learning sections separately to compare with the model of the traditional unused scheme. Two groups of experiments are carried out, wherein the first group respectively generates user item vectors through network representation learning by only adding a user item interaction network, and then the user item vectors are fused through a graph neural network. The second group generates a user vector representation by adding only the user social networks. The two added network representations learning is shown to increase prediction accuracy by the following experimental data table.
Figure BDA0002890781840000132
In the setting of parameters, the parameters included in the model include path length, window size, learning rate and embedding dimension. The following experiment comparison is respectively carried out on the parameters, the path length is 10 and 20, the corresponding window size is 5 and 10, the comparison experiment is carried out by adjusting different embedding dimensions, the learning rate and other parameters, and the experiment result is as follows:
as shown in fig. 4, under the condition that the path length is 10 and the window size is 5, two index values of MAE and RMSE are plotted against the learning rate. Fig. 5 is a graph showing the variation of two index values of MAE and RMSE with the learning rate under the condition that the path length is 20 and the window size is 10. According to experimental results, the scheme has the best effect when the path length is 20, the window size is 10 and the learning rate is 0.004. Fig. 6 and 7 show that when the embedding dimension is selected to be 64 dimensions, the model effect of the scheme is the best, as can be seen by comparing the ciao data set embedding dimension experiment index RMSE and the ciao data set embedding dimension experiment index MAE.
The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and decorations can be made without departing from the spirit of the present invention, and these modifications and decorations should also be regarded as being within the scope of the present invention.

Claims (3)

1. The system recommendation method is characterized by comprising the following specific processing steps:
101) data relation association step: from U ═ U1,u2,...,unDenotes a user data set, V ═ V1,v2,...,vmThe representation of the data set of the goods,
Figure FDA0003662503020000011
representing meta-paths between data, a representing a user or a commodity, R representing a social relationship or a buy and sell relationship; the meta path is generated as follows:
Figure FDA0003662503020000012
wherein n istIs the current node, nt+1In order to be the next node, the node is,
Figure FDA0003662503020000013
among the neighbors representing v are those belonging to At+1ρ represents a meta path rule, and the length of the meta path is set to be L; phi (x) represents the type of the next node of the path, belonging to the node type in the network; v and x respectively represent a current node and a next node, and belong to all nodes in the network;
the vector representation of the user and the commodity generated by the meta path of the interaction relation of the user and the commodity is
Figure FDA0003662503020000014
And
Figure FDA0003662503020000015
the user vector generated by the meta path of the user social relationship is represented as
Figure FDA0003662503020000016
Score vector representation er
102) And (3) user relationship data fusion: the user data set comprises u _ v of a user item interactive network and u _ u of a user social network; the u _ v of the user item interactive network is combined with the attribute network to generate corresponding weight to obtain the implicit vector representation of the user in the user interactive network, and the implicit vector representation is specifically represented by the following formula:
Figure FDA0003662503020000017
wherein,
Figure FDA0003662503020000018
representing the implicit vector of the user i in the interactive network, sigma being an activation function, w and b being parameters of the neural network, C representing all the scores made by the user on the commodity, and NiDenotes the neighbors of user i, softmax denotes the softmax function, xjrVector fusion, g, representing item j and corresponding score rτIs a multi-layer fully-connected neural network,
Figure FDA0003662503020000019
splicing of the representation vectors, w1、w2、b1、b2Representing parameters of two fully connected layers of an attention network;
generating corresponding weight by combining u _ u of the user social network with the attention network to obtain implicit vector representation of the user in the social network, wherein the implicit vector representation is specifically represented by the following formula:
Figure FDA0003662503020000021
wherein,
Figure FDA0003662503020000022
representing an implicit vector of a user i in the social network, and i' representing a neighbor user;
the user vector formula that ultimately includes the user project interaction network and the user social network is as follows:
Figure FDA0003662503020000023
hirepresenting the end-user vector,/n-1Indicating the number of network layers, wn、bnParameters representing w and b multilayer neural networks;
103) and (3) commodity data fusion: the deep relationship of the commodity in the network is collected in a meta-path mode, and the feature vector of the commodity is represented by fusing the features of the neighbor users through GNN
Figure FDA0003662503020000024
The specific formula is as follows:
Figure FDA0003662503020000025
wherein,
Figure FDA0003662503020000026
representing an implicit vector of the commodity j in the interactive network;
104) a recommendation score predicting step: predicting the scores of the user on the commodities according to the user vectors and the commodity vectors obtained in the steps 101) to 103), splicing the user vectors and the commodity vectors through a full-connection layer network to serve as the input of the network, and using the predicted scores as the output of the network, namely using a multilayer full-connection neural network to obtain the final predicted scores, wherein the specific formula is as follows:
Figure FDA0003662503020000031
wherein,
Figure FDA0003662503020000032
is a prediction score.
2. The system recommendation method of claim 1, wherein: the specific process of implicit vector representation of the user in the user interaction network in step 102) is as follows:
Figure FDA0003662503020000033
where aggregate represents the fusion function, xjrThe vectors representing the commodities j and the corresponding scores r are fused, and C represents all scores which are made by the user on the commodities; x is the number ofjrThe formula of (1) is as follows:
Figure FDA0003662503020000034
gτis a multi-layer fully-connected neural network,
Figure FDA0003662503020000035
representing the concatenation of the vectors;
the Aggregation fusion function adopts a fully-connected neural network, and the concrete formula is as follows:
Figure FDA0003662503020000036
σ is the activation function, w and b are neural network parameters, αjiFor the weight of a commodity j to a user i, the importance degree of each commodity to the user is different, the weight is trained through an Attention network, and the specific formula is as follows:
Figure FDA0003662503020000037
3. the system recommendation method of claim 1, wherein: further comprising 105) a parameter adjusting step: adjusting parameters of the recommendation model established in the steps 101) to 104), and specifically, defining an objective function to optimize so as to reversely update the parameters in the recommendation model; the Adam method is selected for adjustment and optimization, wherein the loss formula of adjustment is as follows:
Figure FDA0003662503020000041
wherein,
Figure FDA0003662503020000042
predictive score, y, for user i on item jijFor a true score, n + m is the total number of users and goods.
CN202110027315.5A 2021-01-09 2021-01-09 System recommendation method Active CN112686736B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110027315.5A CN112686736B (en) 2021-01-09 2021-01-09 System recommendation method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110027315.5A CN112686736B (en) 2021-01-09 2021-01-09 System recommendation method

Publications (2)

Publication Number Publication Date
CN112686736A CN112686736A (en) 2021-04-20
CN112686736B true CN112686736B (en) 2022-07-05

Family

ID=75456885

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110027315.5A Active CN112686736B (en) 2021-01-09 2021-01-09 System recommendation method

Country Status (1)

Country Link
CN (1) CN112686736B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104573103A (en) * 2015-01-30 2015-04-29 福州大学 Coauthor recommending method under scientific and technical literature heterogeneous network
CN109871504A (en) * 2019-01-24 2019-06-11 中国科学院软件研究所 A kind of Course Recommendation System based on Heterogeneous Information network and deep learning
CN110633422A (en) * 2019-09-16 2019-12-31 安徽大学 Microblog friend recommendation method based on heterogeneous information network

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107944629B (en) * 2017-11-30 2020-08-07 北京邮电大学 Recommendation method and device based on heterogeneous information network representation
CN109002488B (en) * 2018-06-26 2020-10-02 北京邮电大学 Recommendation model training method and device based on meta-path context
CN111191081B (en) * 2019-12-17 2022-02-22 安徽大学 Developer recommendation method and device based on heterogeneous information network

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104573103A (en) * 2015-01-30 2015-04-29 福州大学 Coauthor recommending method under scientific and technical literature heterogeneous network
CN109871504A (en) * 2019-01-24 2019-06-11 中国科学院软件研究所 A kind of Course Recommendation System based on Heterogeneous Information network and deep learning
CN110633422A (en) * 2019-09-16 2019-12-31 安徽大学 Microblog friend recommendation method based on heterogeneous information network

Also Published As

Publication number Publication date
CN112686736A (en) 2021-04-20

Similar Documents

Publication Publication Date Title
CN111523047B (en) Multi-relation collaborative filtering algorithm based on graph neural network
CN111563164B (en) Specific target emotion classification method based on graph neural network
CN112989064B (en) Recommendation method for aggregating knowledge graph neural network and self-adaptive attention
CN112232925A (en) Method for carrying out personalized recommendation on commodities by fusing knowledge maps
Qu et al. An end-to-end neighborhood-based interaction model for knowledge-enhanced recommendation
CN112950324B (en) Knowledge graph assisted pairwise sorting personalized merchant recommendation method and system
CN109190030B (en) Implicit feedback recommendation method fusing node2vec and deep neural network
CN103971161B (en) Hybrid recommendation method based on Cauchy distribution quantum-behaved particle swarm optimization
Guo et al. Trust-aware recommendation based on heterogeneous multi-relational graphs fusion
Duma et al. Sparseness reduction in collaborative filtering using a nearest neighbour artificial immune system with genetic algorithms
AU2020101604A4 (en) A Recommendation with Item Cooccurrence based on Metric Factorization
CN111881363A (en) Recommendation method based on graph interaction network
CN108470075A (en) A kind of socialization recommendation method of sequencing-oriented prediction
CN109710835A (en) Heterogeneous information network recommendation method with time weight
CN115374347A (en) Social recommendation method based on knowledge graph attention network
CN113051468A (en) Movie recommendation method and system based on knowledge graph and reinforcement learning
CN115329215A (en) Recommendation method and system based on self-adaptive dynamic knowledge graph in heterogeneous network
CN114817712A (en) Project recommendation method based on multitask learning and knowledge graph enhancement
CN117495481A (en) Article recommendation method based on heterogeneous timing diagram attention network
CN115840853A (en) Course recommendation system based on knowledge graph and attention network
CN113342994B (en) Recommendation system based on non-sampling cooperative knowledge graph network
CN114385804A (en) Comment recommendation method of heterogeneous graph attention neural network based on meta-learning
Zhang et al. Knowledge graph driven recommendation model of graph neural network
Chen et al. Graph enhanced neural interaction model for recommendation
CN114842247B (en) Characteristic accumulation-based graph convolution network semi-supervised node classification method

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