CN114092181A - Graph neural network session recommendation method based on disentanglement representation learning - Google Patents

Graph neural network session recommendation method based on disentanglement representation learning Download PDF

Info

Publication number
CN114092181A
CN114092181A CN202111331082.4A CN202111331082A CN114092181A CN 114092181 A CN114092181 A CN 114092181A CN 202111331082 A CN202111331082 A CN 202111331082A CN 114092181 A CN114092181 A CN 114092181A
Authority
CN
China
Prior art keywords
session
commodity
representation
formula
matrix
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
Application number
CN202111331082.4A
Other languages
Chinese (zh)
Inventor
程志勇
李岸松
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.)
Shandong Institute of Artificial Intelligence
Original Assignee
Shandong Institute of Artificial Intelligence
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 Shandong Institute of Artificial Intelligence filed Critical Shandong Institute of Artificial Intelligence
Priority to CN202111331082.4A priority Critical patent/CN114092181A/en
Publication of CN114092181A publication Critical patent/CN114092181A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations
    • 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
    • G06N3/048—Activation functions
    • 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
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Development Economics (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

A graph neural network session recommendation method based on de-entanglement representation learning solves the problems that potential entanglement characteristics cannot be extracted and noise is easily influenced in the session field by a de-entanglement representation learning mechanism and a graph neural network; meanwhile, an attention feature fusion mechanism is provided, so that the problem of excessive smoothness in a graph neural network is relieved; finally, a session de-entanglement representation based on the candidate commodity is provided, so that the representation of the session is richer, and the success of the three points greatly improves the model effect compared with other models in the field.

Description

Graph neural network session recommendation method based on disentanglement representation learning
Technical Field
The invention relates to the technical field of recommendation systems and deep learning, in particular to a graph neural network session recommendation method based on disentanglement representation learning.
Background
In recent years, the recommendation algorithm greatly improves the experience of people in shopping, social contact and the like, so that the recommendation algorithm becomes a main technical direction for the development of each platform. At present, many methods are personalized recommendation based on description of user preferences, but due to the current national control on privacy, a platform cannot easily obtain user information. The recommendation method based on session is applied, and the recommendation method can complete the recommendation of the next commodity only based on an anonymous session, wherein the session is composed of behaviors of the user in a period of time. In this field, the current graph neural network has achieved good results, but some problems are not solved according to the past. One of them is that there is no way to overcome the problem of noise in session, such as users clicking some irrelevant items at will; and the excessive smoothness problem of the graph neural network due to the excessive iteration number. Besides, most importantly, an interpretable process for extracting and modeling a plurality of features cannot be provided, so that feature information acquired by the model is limited, and an optimal model effect cannot be achieved.
Disclosure of Invention
In order to overcome the defects of the technology, the invention provides a method for extracting and modeling a plurality of characteristics of a commodity based on a graph neural network of disentanglement representation.
The technical scheme adopted by the invention for overcoming the technical problems is as follows:
a graph neural network session recommendation method based on disentanglement representation learning comprises the following steps:
a) extracting a session sequence V ═ { V ═ V) containing no user information from a commodity transaction data set1,v2,...vi,...,vnV, the session sequence is composed of several commodities interacting with the user, viSelecting N% of data from the session sequence as training data and the rest 1-N% of data as test data for the ith commodity interacted by the user in the session, wherein i belongs to { 1.. multidot.n }, and N is the number of commodities interacted by the user in the session, and the rest 1-N% of data is used as test data to establish a candidate commodity set L ═ L1,l2,...lj,...,lm}，ljJ belongs to { 1.,. m }, and m is the number of commodities in the session;
b) constructing a session directed graph;
c) establishing and training a graph neural network model based on the disentanglement representation;
d) and calculating the score of the candidate commodity as the next possible interaction object of the session to complete the recommendation of the next commodity. Preferably, in step a), N has a value of 80.
Further, in the above-mentioned case,constructing a session graph G in the step b) by a formula G ═ V, E, wherein E is an edge set of the directed graph, and E ∈ { E } Ei-1,ei}，ei-1,eiIndicates the i-1 st commodity vi-1To the ith commodity viHas a directed edge.
Further, step c) comprises the steps of:
c-1) by the formula
Figure BDA0003348520830000021
Calculating to obtain a commodity v with session embedded through initializationiMapping to the kth feature subspace ci,kIn the formula es,iIs a commodity viInitial embedding of es,i∈RKR is a real number matrix, K is the number of the feature space, | · (| computationally |)2Is L2 norm, WkAs a weight matrix, bkIs a weight matrix, T is a transpose,
Figure BDA0003348520830000022
d is a commercial product viThe original dimension of the beam of light,
Figure BDA0003348520830000023
is the dimension mapped to the kth space;
c-2) by the formula
Figure BDA0003348520830000024
Two neighbor commodities v in k space are obtained through calculationi-1And viIs given a similarity score of
Figure BDA0003348520830000025
ci-1Similarity scores for the feature representation of the i-1 st merchandise in session
Figure BDA0003348520830000026
Constructing an adjacency matrix A together with the session directed graph in step b) as weights between neighborsk,iThe adjacent matrix is formed by splicing an out-degree matrix and an in-degree matrix, and elements in the adjacent matrix are weights between adjacent commodities;
c-3) by the formula
Figure BDA0003348520830000027
Figure BDA0003348520830000028
Figure BDA0003348520830000029
Calculating to obtain the node representation of the ith commodity in the kth feature space,
Figure BDA0003348520830000031
obtaining the related information of the neighbor nodes for the ith commodity in the kth iteration of the kth feature space
Figure BDA0003348520830000032
Wo,Wz,Wr,H,Uo,Uz,UrAll are weight matrices, b is a bias matrix, tanh is a hyperbolic tangent activation function,
Figure BDA0003348520830000033
in order to update the door,
Figure BDA0003348520830000034
to forget gate, σ is sigmoid function,
Figure BDA0003348520830000035
the kth feature subspace c for the t-1 th iterationi,k，
Figure BDA0003348520830000036
The kth feature subspace c for the t-th iterationi,k，
Figure BDA0003348520830000037
To obtain the information before the current round needs to be forgotten according to the forgetting door mechanism
Figure BDA0003348520830000038
Representing the ith point in the kth feature subspace in the session graph of the t-1 iteration, wherein i belongs to { 1.,. n }, and n is the number of points in the session graph;
c-4) obtaining the representation of the ith commodity in the session in each feature space after t rounds of graph neural network iteration
Figure BDA0003348520830000039
Splicing the representations of the ith commodity in each feature space from 1 to k in sequence to form an disentangled representation of the ith commodity
Figure BDA00033485208300000310
c-5) by the formula
Figure BDA00033485208300000311
Calculating to obtain final representation of the commodity
Figure BDA00033485208300000312
Wf、Wq、WpAre all weight matrices, es,iIs a representation of the ith good that did not pass through the neural network; c-6) according to the formula
Figure BDA00033485208300000313
Obtaining the current interest representation of the ith commodity in the session in the kth feature space
Figure BDA00033485208300000314
Is significant coefficient of
Figure BDA00033485208300000315
In the formula W1、W2And p are both a matrix of weights,
Figure BDA00033485208300000316
the ith commodity in session is the kth commodityFeature representation of feature space by formula
Figure BDA00033485208300000317
Calculating the global interest expression of the kth characteristic space of the session
Figure BDA00033485208300000318
By the formula
Figure BDA00033485208300000319
Splicing the global interest representation and the local interest representation to be converted into a session representation in the k characteristic space finally
Figure BDA00033485208300000320
Where | | | is the splicing operation, W3Performing the above operations on K feature spaces to obtain a weight matrix
Figure BDA00033485208300000321
k∈{1,…,K}；
c-7) substituting the candidate commodity set L for the formula in the step c-1)
Figure BDA0003348520830000041
E in (a)s,iMapping the candidate commodity set L to different feature subspaces by using the formula to obtain e ═ c1,…,ck]，ckFor the expression of the candidate commodity in the k characteristic space, the formula is used
Figure BDA0003348520830000042
Calculating to obtain the average representation gamma of the candidate commodity in each feature space according to a formula thetak＝qTσ(W1γ+W2ck+ b) calculating the attention fraction theta of the candidate commodity to the kth feature spacekBy the formula
Figure BDA0003348520830000043
Calculating to obtain the session representation S of the candidate commoditytIn the formula, the | | | is splicing operation and passes through the formula Sh＝W3(St) Calculating to obtain final disentanglement representation S of sessionhIn the formula W3Is a weight matrix;
c-8) by the formula
Figure BDA0003348520830000044
Calculating a likelihood score Z for the candidate good for the next interactioniCompleting the establishment of the neural network model;
c-9) by the formula La＝Lc+λLdecCalculating a loss function LaIn the formula LcThe difference from the true commodity label is calculated for the cross entropy loss function,
Figure BDA0003348520830000045
Ldecthe redundancy of the respective feature subspaces is eliminated as a regularization for the distance correlation function,
Figure BDA0003348520830000046
Figure BDA0003348520830000047
yiin order to be a real label, the label,
Figure BDA0003348520830000048
for the label obtained by the calculation of the graph neural network model, lambda is the coefficient controlling the regular term, dCov is the distance covariance between two matrixes, dVar is the covariance of the matrix itself,
Figure BDA0003348520830000049
for the initial representation matrix of all the commodities in session in the kth feature space,
Figure BDA00033485208300000410
combining the loss function L through BPTT back propagation for the initial representation matrix of all commodities in the (k + 1) th feature space in sessionaOptimizing the graph neural network model of steps c-1) to c-8).
Further, in step d)Test data in the ession sequence are input into a trained graph neural network model to obtain a session de-entanglement representation ShAnd candidate item representation ciBy the formula
Figure BDA0003348520830000051
Calculating a likelihood score Z for the next recommendation for the itemiAccording to the probability score ZiAnd sorting, and recommending the commodity with the highest sorting as the next commodity.
Preferably, in the step c-2), before the splicing operation of the outgoing degree matrix and the incoming degree matrix, each row of the outgoing degree matrix and each row of the incoming degree matrix are respectively standardized.
The invention has the beneficial effects that: through a learning mechanism of disentanglement representation, the problem that potential entanglement characteristics cannot be extracted and are easily influenced by noise in the session field is solved by using a graph neural network; meanwhile, an attention feature fusion mechanism is provided, so that the problem of excessive smoothness in a graph neural network is relieved; finally, a session de-entanglement representation based on the candidate commodity is provided, so that the representation of the session is richer; the success of the three points enables the model effect to be greatly improved compared with other models in the field.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Detailed Description
The invention is further described below with reference to fig. 1.
A graph neural network session recommendation method based on disentanglement representation learning comprises the following steps:
a) extracting a session sequence V ═ { V ═ V) containing no user information from a commodity transaction data set1,v2,…vi,…,vnV, the session sequence is composed of several commodities interacting with the user, viFor the ith commodity interacted by the user in the session, i belongs to {1, …, N }, N is the number of commodities interacted by the user in the session, N% of data is selected from the session sequence as training data, the rest 1-N% of data is used as test data, and a candidate commodity set is establishedL＝{l1,l2,…lj,…,lm}，ljJ belongs to {1, …, m } for the jth commodity in the session, and m is the number of commodities in the session;
b) constructing a session directed graph;
c) establishing and training a graph neural network model based on the disentanglement representation;
d) and calculating the score of the candidate commodity as the next possible interaction object of the session to complete the recommendation of the next commodity. Through a learning mechanism of disentanglement representation, the problem that potential entanglement characteristics cannot be extracted and are easily influenced by noise in the session field is solved by using a graph neural network; meanwhile, an attention feature fusion mechanism is provided, so that the problem of excessive smoothness in a graph neural network is relieved; finally, a session de-entanglement representation based on the candidate commodity is provided, so that the representation of the session is richer; the success of the three points enables the model effect to be greatly improved compared with other models in the field.
Example 1:
in step a), the value of N is 80.
Example 2:
constructing a session graph G in the step b) by a formula G (V, E), wherein V comprises all points (one repeating point is calculated) appearing in V, E is an edge set of a directed graph, and E belongs to { E ∈ { E }i-1,ei}，ei-1,eiIndicates the i-1 st commodity vi-1To the ith commodity viHas a directed edge, i.e. the user has clicked on the commodity vi-1Then click on the commodity vi。
Example 3:
the step c) comprises the following steps:
c-1) by the formula
Figure BDA0003348520830000061
Calculating to obtain a commodity v with session embedded through initializationiMapping to the kth feature subspace ci,kIn the formula es,iIs a commodity viInitial embedding of es,i∈RKR is a real number matrix, K is the number of the feature space, | · (| computationally |)2Is L2 norm, WkAs a weight matrix, bkIs a weight matrix, T is a transpose,
Figure BDA0003348520830000062
d is a commercial product viThe original dimension of the beam of light,
Figure BDA0003348520830000063
is the dimension mapped to the kth space;
c-2) obtaining the representation of each characteristic subspace of the commodity in the session, and then using a formula
Figure BDA0003348520830000064
Two neighbor commodities v in k space are obtained through calculationi-1And viIs given a similarity score of
Figure BDA0003348520830000065
ci-1Similarity scores for the feature representation of the i-1 st merchandise in session
Figure BDA0003348520830000066
Constructing an adjacency matrix A together with the session directed graph in step b) as weights between neighborsk,iThe adjacent matrix is formed by splicing an out-degree matrix and an in-degree matrix, and elements in the adjacent matrix are weights between adjacent commodities;
c-3) by the formula
Figure BDA0003348520830000071
Figure BDA0003348520830000072
Figure BDA0003348520830000073
Calculating to obtain a node table of the ith commodity in the kth feature spaceAs shown in the figure, the material of the steel wire,
Figure BDA0003348520830000074
obtaining the related information of the neighbor nodes for the ith commodity in the kth iteration of the kth feature space
Figure BDA0003348520830000075
Wo,Wz,Wr,H,Uo,Uz,UrAll are weight matrices, b is a bias matrix, tanh is a hyperbolic tangent activation function,
Figure BDA0003348520830000076
in order to update the door,
Figure BDA0003348520830000077
for forgetting, the new information and the old information are considered to calculate the proportion of the new information and the old information, wherein the new information refers to the neighbor information obtained in the round, the old information refers to the representation of the current commodity node, sigma is a sigmoid function,
Figure BDA0003348520830000078
the kth feature subspace c for the t-1 th iterationi,k，
Figure BDA0003348520830000079
The kth feature subspace c for the t-th iterationi,k，
Figure BDA00033485208300000710
To obtain the information before the current round needs to be forgotten according to the forgetting door mechanism
Figure BDA00033485208300000711
Representing the ith point in the kth feature subspace in the session graph of the t-1 iteration, wherein i belongs to { 1.,. n }, and n is the number of points in the session graph;
c-4) obtaining the representation of the ith commodity in the session in each feature space after t rounds of graph neural network iteration
Figure BDA00033485208300000712
Splicing the representations of the ith commodity in each feature space from 1 to k in sequence to form an disentangled representation of the ith commodity
Figure BDA00033485208300000713
c-5) by the formula
Figure BDA00033485208300000714
Calculating to obtain final representation of the commodity
Figure BDA00033485208300000715
Wf、Wq、WpAre all weight matrices, es,iIs a representation of the ith good that did not pass through the neural network; c-6) after obtaining the de-entanglement characteristic representation of the commodities in the session, the general method is to construct the complete session representation only according to the commodities, but in this way, the effect of candidate commodities (commodities to be recommended) is ignored, and we intend to customize a unique session de-entanglement representation for each candidate commodity, firstly according to a formula
Figure BDA00033485208300000716
Obtaining the current interest representation of the ith commodity in the session in the kth feature space
Figure BDA0003348520830000081
Is significant coefficient of
Figure BDA0003348520830000082
In the formula W1、W2And p are both a matrix of weights,
Figure BDA0003348520830000083
for the feature representation of the ith commodity in the kth feature space in the session, the formula is used
Figure BDA0003348520830000084
Calculating the global interest expression of the kth characteristic space of the session
Figure BDA0003348520830000085
By the formula
Figure BDA0003348520830000086
Splicing the global interest representation and the local interest representation to be converted into a session representation in the k characteristic space finally
Figure BDA0003348520830000087
Where | | | is the splicing operation, W3Performing the above operations on K feature spaces to obtain a weight matrix
Figure BDA0003348520830000088
k∈{1,...,K}；
c-7) after obtaining the global interest and the representation of the candidate commodity in each feature subspace, replacing the formula of the step c-1) with the candidate commodity set L
Figure BDA0003348520830000089
E in (a)s,iMapping the candidate commodity set L to different feature subspaces by using the formula to obtain e ═ c1,…,ck]，ckFor the expression of the candidate commodity in the k characteristic space, the formula is used
Figure BDA00033485208300000810
Calculating to obtain the average representation gamma of the candidate commodity in each feature space according to a formula thetak＝qTσ(W1γ+W2ck+ b) calculating the attention fraction theta of the candidate commodity to the kth feature spacekBy the formula
Figure BDA00033485208300000811
Calculating to obtain the session representation S of the candidate commoditytIn the formula, the | | | is splicing operation and passes through the formula Sh＝W3(St) Calculating to obtain final disentanglement representation S of sessionhIn the formula W3Is a weight matrix;
c-8) by the formula
Figure BDA00033485208300000812
Calculating a likelihood score Z for the candidate good for the next interactioniCompleting the establishment of the neural network model;
c-9) by the formula La＝Lc+λLdecCalculating a loss function LaIn the formula LcThe difference from the true commodity label is calculated for the cross entropy loss function,
Figure BDA00033485208300000813
Ldecthe redundancy of the respective feature subspaces is eliminated as a regularization for the distance correlation function,
Figure BDA0003348520830000091
Figure BDA0003348520830000092
yiin order to be a real label, the label,
Figure BDA0003348520830000093
for the label obtained by the calculation of the graph neural network model, lambda is the coefficient controlling the regular term, dCov is the distance covariance between two matrixes, dVar is the covariance of the matrix itself,
Figure BDA0003348520830000094
for the initial representation matrix of all the commodities in session in the kth feature space,
Figure BDA0003348520830000095
combining the loss function L through BPTT back propagation for the initial representation matrix of all commodities in the (k + 1) th feature space in sessionaOptimizing the graph neural network model of steps c-1) to c-8), 0 indicating that the graph neural network has not been trained yet.
Example 4:
step d), inputting the test data in the ession sequence into the trained neural network model to obtainsession de-entanglement representation ShAnd candidate item representation ciBy the formula
Figure BDA0003348520830000096
Calculating a likelihood score Z for the next recommendation for the itemiAccording to the probability score ZiAnd sorting, wherein the score represents the possibility that the commodity is taken as the next interactive object in the current session, and the commodity with the highest sorting is taken as the next commodity for recommendation.
Example 5:
in the step c-2), before splicing operation of the outgoing matrix and the incoming matrix, standardization processing is respectively carried out on each row of the outgoing matrix and the incoming matrix.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (6)

1. A graph neural network session recommendation method based on disentanglement representation learning is characterized by comprising the following steps:
a) extracting a session sequence V ═ { V ═ V) containing no user information from a commodity transaction data set1,v2,...vi,...,vnV, the session sequence is composed of several commodities interacting with the user, viSelecting N% of data from the session sequence as training data, and the rest 1-N% of data as test data to establish a candidate commodity set L ═ L ═ N ∈ {1, …, N } for the ith commodity interacted by the user in the session, wherein N is the number of commodities interacted by the user in the session, and the data of the N% are selected as training data and the rest 1-N% are selected as test data1,l2,…lj,…,lm}，ljFor the jth commodity in session, j is equal to {1, …, m }, and m isThe number of commodities in session;
b) constructing a session directed graph;
c) establishing and training a graph neural network model based on the disentanglement representation;
d) and calculating the score of the candidate commodity as the next possible interaction object of the session to complete the recommendation of the next commodity.
2. The graph neural network session recommendation method based on disentanglement representation learning, according to claim 1, wherein: in step a), the value of N is 80.
3. The graph neural network session recommendation method based on disentanglement representation learning, according to claim 1, wherein: constructing a session graph G in the step b) by a formula G ═ V, E, wherein E is an edge set of the directed graph, and E ∈ { E } Ei-1,ei}，ei-1,eiIndicates the i-1 st commodity vi-1To the ith commodity viHas a directed edge.
4. The disentanglement representation learning-based graph neural network session recommendation method according to claim 1, wherein step c) includes the steps of:
c-1) by the formula
Figure FDA0003348520820000011
Calculating to obtain a commodity v with session embedded through initializationiMapping to the kth feature subspace ci,kIn the formula es,iIs a commodity viInitial embedding of es,i∈RKR is a real number matrix, K is the number of the feature space, | · (| computationally |)2Is L2 norm, WkAs a weight matrix, bkIs a weight matrix, T is a transpose,
Figure FDA0003348520820000012
d is a commercial product viThe original dimension of the beam of light,
Figure FDA0003348520820000013
is the dimension mapped to the kth space;
c-2) by the formula
Figure FDA0003348520820000021
Two neighbor commodities v in k space are obtained through calculationi-1And viIs given a similarity score of
Figure FDA0003348520820000022
ci-1Similarity scores for the feature representation of the i-1 st merchandise in session
Figure FDA0003348520820000023
Constructing an adjacency matrix A together with the session directed graph in step b) as weights between neighborsk,iThe adjacent matrix is formed by splicing an out-degree matrix and an in-degree matrix, and elements in the adjacent matrix are weights between adjacent commodities;
c-3) by the formula
Figure FDA0003348520820000024
Figure FDA0003348520820000025
Figure FDA0003348520820000026
Calculating to obtain the node representation of the ith commodity in the kth feature space,
Figure FDA0003348520820000027
obtaining the related information of the neighbor nodes for the ith commodity in the kth iteration of the kth feature space
Figure FDA0003348520820000028
Wo,Wz,Wr,H,Uo,Uz,UrAll are weight matrices, b is a bias matrix, tanh is a hyperbolic tangent activation function,
Figure FDA0003348520820000029
in order to update the door,
Figure FDA00033485208200000210
to forget gate, σ is sigmoid function,
Figure FDA00033485208200000211
the kth feature subspace c for the t-1 th iterationi,k，
Figure FDA00033485208200000212
The kth feature subspace c for the t-th iterationi,k，
Figure FDA00033485208200000213
To obtain the information before the current round needs to be forgotten according to the forgetting door mechanism
Figure FDA00033485208200000214
Figure FDA00033485208200000215
Representing the ith point in the kth feature subspace in the session graph of the t-1 th iteration, wherein i belongs to {1, …, n }, and n is the number of points in the session graph;
c-4) obtaining the representation of the ith commodity in the session in each feature space after t rounds of graph neural network iteration
Figure FDA00033485208200000216
Splicing the representations of the ith commodity in each feature space from 1 to k in sequence to form an disentangled representation of the ith commodity
Figure FDA00033485208200000217
c-5) by the formula
Figure FDA00033485208200000218
Calculating to obtain final representation of the commodity
Figure FDA00033485208200000219
Wf、Wq、WpAre all weight matrices, es,iIs a representation of the ith good that did not pass through the neural network;
c-6) according to the formula
Figure FDA0003348520820000031
Obtaining the current interest representation of the ith commodity in the session in the kth feature space
Figure FDA0003348520820000032
Is significant coefficient of
Figure FDA0003348520820000033
In the formula W1、W2And p are both a matrix of weights,
Figure FDA0003348520820000034
for the feature representation of the ith commodity in the kth feature space in the session, the formula is used
Figure FDA0003348520820000035
Calculating the global interest expression of the kth characteristic space of the session
Figure FDA0003348520820000036
By the formula
Figure FDA0003348520820000037
Representing global interests with local interestsThe representation concatenation is converted into a session representation finally in the kth feature space
Figure FDA0003348520820000038
Where | | | is the splicing operation, W3Performing the above operations on K feature spaces to obtain a weight matrix
Figure FDA0003348520820000039
c-7) substituting the candidate commodity set L for the formula in the step c-1)
Figure FDA00033485208200000310
E in (a)s,iMapping the candidate commodity set L to different feature subspaces by using the formula to obtain e ═ c1,...,ck]，ckFor the expression of the candidate commodity in the k characteristic space, the formula is used
Figure FDA00033485208200000311
Calculating to obtain the average representation gamma of the candidate commodity in each feature space according to a formula thetak＝qTσ(W1γ+W2ck+ b) calculating the attention fraction theta of the candidate commodity to the kth feature spacekBy the formula
Figure FDA00033485208200000312
Calculating to obtain the session representation S of the candidate commoditytIn the formula, the | | | is splicing operation and passes through the formula Sh＝W3(St) Calculating to obtain final disentanglement representation S of sessionhIn the formula W3Is a weight matrix;
c-8) by the formula
Figure FDA00033485208200000313
Calculating a likelihood score Z for the candidate good for the next interactioniCompleting the establishment of the neural network model;
c-9) by the formula La＝Lc+λLdecCalculating a loss function LaIn the formula LcThe difference from the true commodity label is calculated for the cross entropy loss function,
Figure FDA00033485208200000314
Ldecthe redundancy of the respective feature subspaces is eliminated as a regularization for the distance correlation function,
Figure FDA0003348520820000041
Figure FDA0003348520820000042
yiin order to be a real label, the label,
Figure FDA0003348520820000043
for the label obtained by the calculation of the graph neural network model, lambda is the coefficient controlling the regular term, dCov is the distance covariance between two matrixes, dVar is the covariance of the matrix itself,
Figure FDA0003348520820000044
for the initial representation matrix of all the commodities in session in the kth feature space,
Figure FDA0003348520820000045
combining the loss function L through BPTT back propagation for the initial representation matrix of all commodities in the (k + 1) th feature space in sessionaOptimizing the graph neural network model of steps c-1) to c-8).
5. The graph neural network session recommendation method based on disentanglement representation learning, according to claim 1, wherein: step d), inputting the test data in the ession sequence into the trained graph neural network model to obtain the session disentanglement representation ShAnd candidate item representation ciBy the formula
Figure FDA0003348520820000046
Calculating a likelihood score Z for the next recommendation for the itemiAccording to the probability score ZiAnd sorting, and recommending the commodity with the highest sorting as the next commodity.
6. The graph neural network session recommendation method based on disentanglement representation learning, according to claim 1, wherein: in the step c-2), before splicing operation of the outgoing matrix and the incoming matrix, standardization processing is respectively carried out on each row of the outgoing matrix and the incoming matrix.
CN202111331082.4A 2021-11-11 2021-11-11 Graph neural network session recommendation method based on disentanglement representation learning Pending CN114092181A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111331082.4A CN114092181A (en) 2021-11-11 2021-11-11 Graph neural network session recommendation method based on disentanglement representation learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111331082.4A CN114092181A (en) 2021-11-11 2021-11-11 Graph neural network session recommendation method based on disentanglement representation learning

Publications (1)

Publication Number Publication Date
CN114092181A true CN114092181A (en) 2022-02-25

Family

ID=80299835

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111331082.4A Pending CN114092181A (en) 2021-11-11 2021-11-11 Graph neural network session recommendation method based on disentanglement representation learning

Country Status (1)

Country Link
CN (1) CN114092181A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114780866A (en) * 2022-04-11 2022-07-22 山东省人工智能研究院 Personalized intelligent recommendation method based on spatio-temporal context interest learning model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114780866A (en) * 2022-04-11 2022-07-22 山东省人工智能研究院 Personalized intelligent recommendation method based on spatio-temporal context interest learning model
CN114780866B (en) * 2022-04-11 2022-11-01 山东省人工智能研究院 Personalized intelligent recommendation method based on spatio-temporal context interest learning model

Similar Documents

Publication Publication Date Title
Sim et al. Is deep learning for image recognition applicable to stock market prediction?
Ha et al. Response models based on bagging neural networks
CN110362738B (en) Deep learning-based individual recommendation method combining trust and influence
CN111949865A (en) Interest point recommendation method based on graph neural network and user long-term and short-term preference
Romdhane et al. An efficient approach for building customer profiles from business data
CN111783474A (en) Comment text viewpoint information processing method and device and storage medium
CN110503531A (en) The dynamic social activity scene recommended method of timing perception
CN114817663B (en) Service modeling and recommendation method based on class perception graph neural network
CN113407864B (en) Group recommendation method based on mixed attention network
CN113269647B (en) Graph-based transaction abnormity associated user detection method
CN112527966B (en) Network text emotion analysis method based on Bi-GRU neural network and self-attention mechanism
CN113297936B (en) Volleyball group behavior identification method based on local graph convolution network
CN113609388B (en) Sequence recommendation method based on anti-facts user behavior sequence generation
CN112700274A (en) Advertisement click rate estimation method based on user preference
CN112906890A (en) User attribute feature selection method based on mutual information and improved genetic algorithm
CN112070543A (en) Method for detecting comment quality in E-commerce website
US20200272812A1 (en) Human body part segmentation with real and synthetic images
CN111400525B (en) Fashion clothing intelligent matching and recommending method based on vision combination relation learning
CN114780866B (en) Personalized intelligent recommendation method based on spatio-temporal context interest learning model
CN114092181A (en) Graph neural network session recommendation method based on disentanglement representation learning
KR20230069578A (en) Sign-Aware Recommendation Apparatus and Method using Graph Neural Network
CN114036298A (en) Node classification method based on graph convolution neural network and word vector
CN112784177B (en) Spatial distance adaptive next interest point recommendation method
CN117408735A (en) Client management method and system based on Internet of things
JP2009503657A (en) Rating system using neural network

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