CN110991601B - Neural network recommendation method based on multi-user behavior - Google Patents

Neural network recommendation method based on multi-user behavior Download PDF

Info

Publication number
CN110991601B
CN110991601B CN201911207316.7A CN201911207316A CN110991601B CN 110991601 B CN110991601 B CN 110991601B CN 201911207316 A CN201911207316 A CN 201911207316A CN 110991601 B CN110991601 B CN 110991601B
Authority
CN
China
Prior art keywords
behavior
user
behaviors
target
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.)
Active
Application number
CN201911207316.7A
Other languages
Chinese (zh)
Other versions
CN110991601A (en
Inventor
吴迪
林诗鹭
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.)
Sun Yat Sen University
Original Assignee
Sun Yat Sen University
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 Sun Yat Sen University filed Critical Sun Yat Sen University
Priority to CN201911207316.7A priority Critical patent/CN110991601B/en
Publication of CN110991601A publication Critical patent/CN110991601A/en
Application granted granted Critical
Publication of CN110991601B publication Critical patent/CN110991601B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides a neural network recommendation method based on multi-user behaviors, which comprises the following steps: s1, mining the relevance between non-target behaviors and target behaviors, and obtaining the confidence coefficient from the non-target behaviors to the target behaviors; s2, a neural network recommendation model based on multiple behaviors of the user is used for mining the relation between the multiple behaviors and the user preferences, wherein the relation comprises the steps of constructing a behavior feature matrix of the user and the object, applying a neural collaborative filtering model, and solving the model by using a behavior classification learning strategy based on the confidence coefficient obtained in the S1 to obtain an object recommendation list. According to the neural network recommendation method based on multi-user behaviors, firstly, user behavior data are divided according to time periods, occurrence probability from non-target behaviors to target behaviors is calculated, behavior feature matrixes of users and articles are constructed according to interactive behaviors of the users and the articles, then a neural collaborative filtering model is utilized to represent complex relations between behaviors and user favorites, and finally, a model is solved by using a strategy of behavior classification learning to obtain a recommendation list.

Description

Neural network recommendation method based on multi-user behavior
Technical Field
The invention relates to the field of recommendation systems, in particular to a neural network recommendation method based on multi-user behaviors.
Background
There are two main types of recommendation methods based on multi-user behavior at present. One is a method of using joint matrix decomposition, in which a user is represented as a plurality of feature matrices according to different behaviors, and interactions between the user and an article are predicted by performing inner products with the feature matrices of the article. One is a method using bayesian personalized ordering, which orders items according to the relative importance of the interaction behavior of a user with a specific item, and generates a recommendation list. The main problem with the first type of approach is that it represents a linear relationship between user behavior and preferences, and it is difficult to abstract more complex nonlinear relationships. The second type of method has the main problem that the relative importance degree of interaction behavior needs to be set manually, and large errors exist.
In recent years, the development of the machine learning field provides a guide for a recommendation system. The neural network can effectively acquire nonlinear relations, a model of a recommendation method can be designed by using the neural network, and deeper and complex relations between user behaviors and preferences can be abstracted. Furthermore, there is a certain correlation between the various behaviors of the user, and the degree of this correlation may change over time, that is, the importance of a particular behavior to reflect its preferences may change. To fully account for this dynamic change, a Markov model is used in our algorithm to capture the dynamic change in the correlation between behaviors and to learn the model using a strategy of behavior classification learning as one of the parameters of the neural network model.
Disclosure of Invention
The recommendation method based on multiple users in the prior art has the following common problems: it is difficult to abstract the complex nonlinear relationship between user behavior and preference, and there is a large error in the recommendation result.
In order to solve the technical problems, the technical scheme of the invention is as follows:
the invention discloses a neural network recommendation system based on multi-user behaviors, which comprises a behavior association algorithm and a neural network recommendation model based on the multi-user behaviors;
the behavior association algorithm is used for mining the association between the target behavior and obtaining the confidence coefficient from the non-target behavior to the target behavior;
the neural network recommendation model based on the user multi-behavior is used for mining the relation between the multi-behavior and the user preference, wherein a behavior feature matrix of the user and the object is constructed, a neural collaborative filtering model is applied, the model is solved by using a behavior classification learning strategy according to the confidence coefficient obtained by a behavior association algorithm, and finally an object recommendation list is obtained.
The invention discloses a neural network recommendation method based on multi-user behaviors, which comprises the following steps:
s1, mining the relevance of non-target behaviors and target behaviors by using a behavior relevance analysis algorithm, and obtaining the confidence coefficient from the non-target behaviors to the target behaviors;
s2, a neural network recommendation model based on multiple behaviors of the user is used for mining the relation between the multiple behaviors and the user preferences, wherein the relation comprises the steps of constructing a behavior feature matrix of the user and the object, applying a neural collaborative filtering model, and solving the model by using a behavior classification learning strategy based on the confidence coefficient obtained in the S1 to obtain an object recommendation list.
In a preferred embodiment, S1 includes the following steps:
s11, using U to represent a user set, I to represent an article set, B to represent a non-target behavior set, O to represent a target behavior set, and using behavior data of the user UThe set is denoted as D u Each set of data in the dataset represents an action that the user has taken place on item i; wherein non-target behavior b j And target behavior o w The relationship between is expressed as an association rule R: b j →o w The confidence of this association is denoted as c (b j →o w ) For representation at b j In the case of occurrence, by association rule "b j →o w "occurrence of push out o w Probability of (2);
s12, association rule b j →o w Confidence sequence C of (2) jw Has the following expression form:
C jw ＝{c(b j →o w ) 1 ,c(b j →o w ) 2 ,…c(b j →o w ) t }
wherein b j For j-th non-target behavior, o w For the w-th target behavior, t is the current time period, c (b j →o w ) t For association rule b j →o w At data subset D ut (
Figure BDA0002297194530000021
Is for data set D u Subsets of data divided by time period).
S13, the confidence coefficient is in the range of 0,1]The interval can be divided into n subintervals (n is a super parameter), and n integers are respectively corresponding to the interval to represent states; obtaining a confidence state sequence S according to the confidence sequence and the conversion rules jw ：
S jw ＝{s(b j →o w ) 1 ,s(b j →o w ) 2 ,…s(b j →o w ) t }；
Wherein b j For j-th non-target behavior, o w For the w-th target behavior, t is the current time period, s (b j →o w ) t For association rule b j →o w At data subset D ut (
Figure BDA0002297194530000022
Is for data set D u Subsets of data divided by time period).
S14, calculating an association rule b by using the confidence state sequence j →o w The transition probability p of each state to any other state xy ＝p(s x →s y ) j Form a state transition matrix P jw ：
Figure BDA0002297194530000031
S15, according to the confidence state sequence obtained in S13, the final state is S (b j →o w ) t The state vector is represented by single hot code
Figure BDA0002297194530000032
The predicted next state vector is: />
Figure BDA0002297194530000033
S16, letting index be
Figure BDA0002297194530000034
Subscript corresponding to maximum value in (b), confidence state s (b) of time period t j →o w ) t+1 Index, i.e. confidence of next state c (b j →o w ) t+1 ＝s(b j →o w ) t+1 Corresponds to the median value of the confidence interval.
In a preferred embodiment, the specific steps at S2 are as follows:
s21, encoding IDs of the user and the article respectively by using independent thermal encoding according to a characteristic matrix P of the user and a characteristic matrix Q of the article, and constructing a behavior characteristic matrix according to the multi-behavior characteristics of the user;
s22, inputting the behavior feature matrix obtained in the step S21, and using a neural collaborative filtering algorithm to represent a learning function of the behavior;
s23, performing independent learning on non-target behaviors and performing multi-behavior learning strategies of target behavior joint learning, and generating a recommendation list for a user according to the result of the joint learning.
In a preferred embodiment, the specific step S21 is as follows:
s211, generating a user characteristic matrix P and an article characteristic matrix Q, and respectively encoding the IDs of the user and the article by using single thermal encoding to obtain M×E and N×E matrixes;
s212, adding K features to the M×E user feature matrix, wherein the K features correspond to the occurrence times of K behaviors of the user; similarly, K features are added for the N x E article feature matrix, the number of times of occurrence of K behaviors related to the article is respectively corresponding to the number of times of occurrence of K behaviors related to the article, and the sizes of the constructed user feature matrix P and the article feature matrix Q are M x (E+K) and N x (E+K) respectively;
S213.
Figure BDA0002297194530000035
and->
Figure BDA0002297194530000036
The unique heat encoded vectors representing user u and item i, respectively, may be obtained using the following formulas, namely, the ith row of the matrix P and the ith row of Q:
Figure BDA0002297194530000037
Figure BDA0002297194530000038
in a preferred embodiment, the output of the neural network in S22 is
Figure BDA0002297194530000039
Wherein the method comprises the steps of
Figure BDA00022971945300000310
For a functional representation with respect to feature vectors, σ is a sigmoid function that converts the output into probabilities.
In a preferred embodiment, the
Figure BDA00022971945300000311
The method comprises 3 commonly used function expression methods, namely a general matrix decomposition method (GMF), a multi-layer perceptron Method (MLP) and a neural matrix decomposition method (NeuMF), and the three methods are described as follows;
the general matrix factorization method (GMF) is to add weights to each term on the result of matrix factorization:
Figure BDA0002297194530000041
where h represents a weight vector.
The multi-layer perceptron Method (MLP) learns by a nonlinear method:
Figure BDA0002297194530000042
…
z L ＝ReLU(W L z L-1 +b L )
Figure BDA0002297194530000043
wherein L represents the number of hidden layers in the multi-layer perceptron, W x And b x The weight and bias items of the x layer are respectively expressed, and the linear rectification function (ReLU) is used as an activation function by default;
the neural matrix decomposition method (NeuMF) combines GMF and MLP according to a certain weight:
Figure BDA0002297194530000044
where h represents a weight vector, z L Is the result of the last layer of the hidden layer obtained through MLP learning.
In a preferred embodiment, the step S23 includes the following steps:
s231, independently learning non-target behaviors;
s2311 user-item is in non-target behavior b j Is represented as a matrix Y j ，Y j The size of (2) is M×N, M is the number of users, and N is the number of articles.
Figure BDA0002297194530000045
S2312 Each non-target behavior b j There is an independent prediction function and loss function. b j The prediction function of (2) is as follows:
Figure BDA0002297194530000046
wherein the method comprises the steps of
Figure BDA0002297194530000047
Is the article in action b j Bias item on->
Figure BDA0002297194530000048
Is a functional representation; />
S2313 for Single action b j Likelihood functions of (2) are as follows:
Figure BDA0002297194530000049
wherein Y is j + Representing a behavior matrix Y j Item 1 in (C), Y j - Representing a behavior matrix Y j An item of 0;
s2314 non-target behavior b j The loss function of (2) is as follows:
Figure BDA0002297194530000051
optimizing the loss function of each non-objective behavioral function independently using a stochastic gradient descent method;
s232, performing joint learning on target behaviors;
s2321, user and article are in target behavior o w Is represented as a matrix Z w ，Z w The size of (2) is M x N, M is the number of users, N is the number of articles,
Figure BDA0002297194530000052
s2322 for target behavior o w For that reason, consider non-target behavior and its relevance.
Figure BDA0002297194530000053
Wherein c jw For the confidence coefficient c (b) obtained in step S16 j →o w ) t+1 ，
Figure BDA0002297194530000054
Is the article in behavior o w The bias term on the upper part of the table,
Figure BDA0002297194530000055
is a functional representation;
s2323 for a single target behavior o w Likelihood functions of (2) are as follows:
Figure BDA0002297194530000056
wherein the method comprises the steps of
Figure BDA0002297194530000057
Representing a behavior matrix Z w Item 1 in>
Figure BDA0002297194530000058
Representing a behavior matrix Z w An item of 0;
s2324. the joint probability function of all non-target behaviors is expressed as:
Figure BDA0002297194530000059
s2325, taking the negative logarithm of the probability function, the following loss function can be obtained, namely the objective function to be optimized:
Figure BDA00022971945300000510
s.t.
Figure BDA00022971945300000511
wherein lambda is w For target action o w The importance of (2) can be obtained through priori knowledge;
s233, constructing a recommendation list according to the prediction results obtained in the S231 and the S232;
predicting the probability of target behavior occurrence according to the previous step
Figure BDA00022971945300000512
Weighting the likelihood of all target behaviors may reflect the preference of user u for item i: />
Figure BDA0002297194530000061
And after the R values of all the articles are ranked from large to small, recommending the articles ranked in front to the user.
Compared with the prior art, the technical scheme of the invention has the beneficial effects that:
the behavior association analysis algorithm provided by the invention is an association analysis algorithm based on a Markov model, the user behavior data is divided according to time periods, and the occurrence probability from non-target behavior to target behavior is calculated. In the multi-behavior neural network recommendation model, the interaction behaviors of the user and the article are considered, a behavior feature matrix of the user and the article is constructed, and the complex nonlinear relation between the user behavior and the preference is represented by utilizing neural collaborative filtering. Solving the model by using a strategy of behavior classification learning: and for the target behavior, the influence of the non-target behavior is considered, and the joint learning is performed.
Drawings
FIG. 1 is a flow chart of the steps of the present invention;
fig. 2 is a schematic diagram of a neural network recommendation model in example 2.
Detailed Description
The drawings are for illustrative purposes only and are not to be construed as limiting the present patent;
for the purpose of better illustrating the embodiments, certain elements of the drawings may be omitted, enlarged or reduced and do not represent the actual product dimensions;
it will be appreciated by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The technical scheme of the invention is further described below with reference to the accompanying drawings and examples.
Example 1
The first aspect of the invention discloses a neural network recommendation system based on multi-user behaviors, as shown in fig. 1, comprising a behavior association algorithm and a neural network recommendation model based on the multi-user behaviors;
the behavior association algorithm is used for mining the association between the target behavior and obtaining the confidence coefficient from the non-target behavior to the target behavior;
the neural network recommendation model based on the user multi-behavior is used for mining the relation between the multi-behavior and the user preference, wherein a behavior feature matrix of the user and the object is constructed, a neural collaborative filtering model is applied, the model is solved by using a behavior classification learning strategy according to the confidence coefficient obtained by a behavior association algorithm, and finally an object recommendation list is obtained.
Example 2
The invention discloses a neural network recommendation method based on multi-user behaviors, which comprises the following steps:
s1, mining the relevance of non-target behaviors and target behaviors by using a behavior relevance analysis algorithm, and obtaining the confidence coefficient from the non-target behaviors to the target behaviors;
s2, as shown in FIG. 2, a neural network recommendation model based on multiple behaviors of a user is used for mining the relation between the multiple behaviors and user preferences, wherein the relation comprises the steps of constructing a behavior feature matrix of the user and the object, applying a neural collaborative filtering model, and solving the model by using a behavior classification learning strategy based on the confidence coefficient obtained in S1 to obtain an object recommendation list.
In a preferred embodiment, S1 includes the following steps:
s11, using U to represent a user set, I to represent an article set, B to represent a non-target behavior set, O to represent a target behavior set, and using the behavior data set of the user U as D u Each set of data in the dataset represents an action that the user has taken place on item i; wherein non-target behavior b j And target behavior o w The relationship between is expressed as an association rule R: b j →o w The confidence of this association is denoted as c (b j →o w ) For representation at b j In the case of occurrence, by association rule "b j →o w "occurrence of push out o w Probability of (2);
s12, association rule b j →o w Confidence sequence C of (2) jw Has the following expression form:
C jw ＝{c(b j →o w ) 1 ,c(b j →o w ) 2 ,…c(b j →o w ) t }
wherein b j For j-th non-target behavior, o w For the w-th target behavior, t is the current time period, c (b j →o w ) t For association rule b j →o w At data subset D ut (
Figure BDA0002297194530000071
Is for data set D u Subsets of data divided by time period).
S13, the confidence coefficient is in the range of 0,1]The interval can be divided into n subintervals (n is a super parameter), and n integers are respectively corresponding to the interval to represent states; obtaining a confidence state sequence S according to the confidence sequence and the conversion rules jw ：
S jw ＝{s(b j →o w ) 1 ,s(b j →o w ) 2 ,…s(b j →o w ) t }；
Wherein b j For j-th non-target behavior, o w For the w-th target behavior, t is the current time period, s (b j →o w ) t For association rule b j →o w At data subset D ut (
Figure BDA0002297194530000072
Is for data set D u Subsets of data divided by time period).
S14, calculating an association rule b by using the confidence state sequence j →o w The transition probability p of each state to any other state xy ＝p(s x →s y ) j Form a state transition matrix P jw ：
Figure BDA0002297194530000073
S15, according to the confidence state sequence obtained in S13, the final state is S (b j →o w ) t The state vector is represented by single hot code
Figure BDA0002297194530000081
The predicted next state vector is: />
Figure BDA0002297194530000082
S16, letting index be
Figure BDA0002297194530000083
Subscript corresponding to maximum value in (b), confidence state s (b) of time period t j →o w ) t+1 Index, i.e. confidence of next state c (b j →o w ) t+1 ＝s(b j →o w ) t+1 Corresponds to the median value of the confidence interval.
In a preferred embodiment, the specific steps at S2 are as follows:
s21, encoding IDs of the user and the article respectively by using independent thermal encoding according to a characteristic matrix P of the user and a characteristic matrix Q of the article, and constructing a behavior characteristic matrix according to the multi-behavior characteristics of the user;
s22, inputting the behavior feature matrix obtained in the step S21, and using a neural collaborative filtering algorithm to represent a learning function of the behavior;
s23, performing independent learning on non-target behaviors and performing multi-behavior learning strategies of target behavior joint learning, and generating a recommendation list for a user according to the result of the joint learning.
In a preferred embodiment, the specific step S21 is as follows:
s211, generating a user characteristic matrix P and an article characteristic matrix Q, and respectively encoding the IDs of the user and the article by using single thermal encoding to obtain M×E and N×E matrixes;
s212, adding K features to the M×E user feature matrix, wherein the K features correspond to the occurrence times of K behaviors of the user; similarly, K features are added for the N x E article feature matrix, the number of times of occurrence of K behaviors related to the article is respectively corresponding to the number of times of occurrence of K behaviors related to the article, and the sizes of the constructed user feature matrix P and the article feature matrix Q are M x (E+K) and N x (E+K) respectively;
S213.
Figure BDA0002297194530000084
and->
Figure BDA0002297194530000085
The unique heat encoded vectors representing user u and item i, respectively, may be obtained using the following formulas, namely, the ith row of the matrix P and the ith row of Q:
Figure BDA0002297194530000086
Figure BDA0002297194530000087
in a preferred embodiment, the output of the neural network in S22 is
Figure BDA0002297194530000088
Wherein the method comprises the steps of
Figure BDA0002297194530000089
For a functional representation with respect to feature vectors, σ is a sigmoid function that converts the output into probabilities.
In a preferred embodiment, the
Figure BDA00022971945300000810
The method comprises 3 commonly used function expression methods, namely a general matrix decomposition method, a multi-layer perceptron method and a neural matrix decomposition method, and the three methods are briefly introduced as follows:
the general matrix decomposition method is to add weights to each term on the result of matrix decomposition:
Figure BDA00022971945300000811
where h represents a weight vector.
The multi-layer perceptron method is to learn by a nonlinear method:
Figure BDA0002297194530000091
…
z L ＝ReLU(W L z L-1 +b L )
Figure BDA0002297194530000092
wherein L represents the number of hidden layers in the multi-layer perceptron, W x And b x The weight and the bias item of the x layer are respectively expressed, and the linear rectification function is used as the activation function by default;
the neural matrix decomposition combines the GMF and the MLP according to a certain weight:
Figure BDA0002297194530000093
where h represents a weight vector, z L Is the result of the last layer of the hidden layer obtained through MLP learning.
In a preferred embodiment, the step S23 includes the following steps:
s231, independently learning non-target behaviors;
s2311 user-item is in non-target behavior b j Is represented as a matrix Y j ，Y j The size of (2) is M×N, M is the number of users, and N is the number of articles.
Figure BDA0002297194530000094
S2312 Each non-target behavior b j There is an independent prediction function and loss function. b j The prediction function of (2) is as follows:
Figure BDA0002297194530000095
wherein the method comprises the steps of
Figure BDA0002297194530000096
Is the article in action b j Bias item on->
Figure BDA0002297194530000097
Is a functional representation;
s2313 for Single action b j Likelihood functions of (2) are as follows:
Figure BDA0002297194530000098
wherein Y is j + Representing a behavior matrix Y j Item 1 in (C), Y j - Representing a behavior matrix Y j An item of 0;
s2314 non-target behavior b j The loss function of (2) is as follows:
Figure BDA0002297194530000099
optimizing the loss function of each non-objective behavioral function independently using a stochastic gradient descent method;
s232, performing joint learning on target behaviors;
s2321, user and article are in target behavior o w Is represented as a matrix Z w ，Z w The size of (2) is M x N, M is the number of users, N is the number of articles,
Figure BDA0002297194530000101
s2322 for target behavior o w For that reason, consider non-target behavior and its relevance.
Figure BDA0002297194530000102
Wherein c jw For the confidence coefficient c (b) obtained in step S16 j →o w ) t+1 ，
Figure BDA0002297194530000103
Is the article in behavior o w The bias term on the upper part of the table,
Figure BDA0002297194530000104
is a functional representation;
s2323 for a single target behavior o w Likelihood functions of (2) are as follows:
Figure BDA0002297194530000105
wherein the method comprises the steps of
Figure BDA0002297194530000106
Representing a behavior matrix Z w Item 1 in>
Figure BDA0002297194530000107
Representing a behavior matrix Z w An item of 0;
s2324. the joint probability function of all non-target behaviors is expressed as:
Figure BDA0002297194530000108
s2325, taking the negative logarithm of the probability function, the following loss function can be obtained, namely the objective function to be optimized:
Figure BDA0002297194530000109
s.t.
Figure BDA00022971945300001010
wherein lambda is w For target action o w The importance of (2) can be obtained through priori knowledge;
s233, constructing a recommendation list according to the prediction results obtained in the S231 and the S232;
predicting the probability of target behavior occurrence according to the previous step
Figure BDA00022971945300001011
Weighting the likelihood of all target behaviors may reflect the preference of user u for item i: />
Figure BDA00022971945300001012
And after the R values of all the articles are ranked from large to small, recommending the articles ranked in front to the user.
The terms describing the positional relationship in the drawings are merely illustrative, and are not to be construed as limiting the present patent;
it is to be understood that the above examples of the present invention are provided by way of illustration only and not by way of limitation of the embodiments of the present invention. Other variations or modifications of the above teachings will be apparent to those of ordinary skill in the art. It is not necessary here nor is it exhaustive of all embodiments. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the invention are desired to be protected by the following claims.

Claims (4)

1. The neural network recommendation method based on multi-user behavior is characterized by comprising the following steps of:
s1, mining the relevance of non-target behaviors and target behaviors by using a behavior relevance analysis algorithm, and obtaining the confidence coefficient from the non-target behaviors to the target behaviors; the method comprises the following specific steps:
s11, using U to represent a user set, I to represent an article set, B to represent a non-target behavior set, O to represent a target behavior set, and using the behavior data set of the user U as D u Data setRepresenting the behaviour of the user on item i; wherein non-target behavior b j And target behavior o w The relationship between is expressed as an association rule R: b j →o w The confidence of this association is denoted as c (b j →o w ) For representation at b j In the case of occurrence, by association rule "b j →o w "occurrence of push out o w Probability of (2);
s12, association rule b j →o w Confidence sequence C of (2) jw Has the following expression form:
C jw ＝{c(b j →o w ) 1 ,c(b j →o w ) 2 ,…c(b j →o w ) t }
wherein b j For j-th non-target behavior, o w For the w-th target behavior, t is the current time period, c (b j →o w ) t For association rule b j →o w At data subset D ut Confidence in (c); wherein the method comprises the steps of
Figure FDA0004104019720000015
Is for data set D u Data subsets divided by time period;
s13, the confidence coefficient is in the range of 0,1]Dividing the interval into n subintervals, wherein n is a super parameter and corresponds to n integers respectively for representing states; obtaining a confidence state sequence S according to the confidence sequence and the conversion rules jw ：
S jw ＝{s(b j →o w ) 1 ,s(b j →o w ) 2 ,…s(b j →o w ) t }；
Wherein s (b) j →o w ) t For association rule b j →o w At data subset D ut Confidence state in (c);
s14, calculating an association rule b by using the confidence state sequence j →o w The transition probability p of each state to any other state xy ＝p(s x →s y ) j Form a state transition matrix P jw ：
Figure FDA0004104019720000011
S15, according to the confidence state sequence obtained in S13, the final state is S (b j →o w ) t The state vector is represented by single hot code
Figure FDA0004104019720000012
The predicted next state vector is: />
Figure FDA0004104019720000013
S16, letting index be
Figure FDA0004104019720000014
Subscript corresponding to maximum value in (b), confidence state s (b) of time period t j →o w ) t+1 Index, i.e. confidence of next state c (b j →o w ) t+1 ＝s(b j →o w ) t+1 A median value of the corresponding confidence intervals;
s2, mining the relation between the multiple behaviors and the user preferences by using a neural network recommendation model based on the multiple behaviors of the user, wherein the method comprises the steps of constructing a behavior feature matrix of the user and the object, applying a neural collaborative filtering model, and solving the model by using a behavior classification learning strategy based on the confidence coefficient obtained in the S1 to obtain an object recommendation list; the method comprises the following specific steps:
s21, encoding IDs of the user and the article respectively by using independent thermal encoding according to a characteristic matrix P of the user and a characteristic matrix Q of the article, and constructing a behavior characteristic matrix according to the multi-behavior characteristics of the user; wherein:
s211, generating a user characteristic matrix P and an article characteristic matrix Q, and respectively encoding the IDs of the user and the article by using single thermal encoding to obtain M×E and N×E matrixes;
s212, adding K features to the M×E user feature matrix, wherein the K features correspond to the occurrence times of K behaviors of the user; similarly, K features are added for the N x E article feature matrix, the number of times of occurrence of K behaviors related to the article is respectively corresponding to the number of times of occurrence of K behaviors related to the article, and the sizes of the constructed user feature matrix P and the article feature matrix Q are M x (E+K) and N x (E+K) respectively;
S213.
Figure FDA0004104019720000027
and->
Figure FDA0004104019720000028
The unique heat coded vectors representing user u and item i, respectively, are used to obtain the corresponding feature vectors, i.e., the ith row of the matrix P and the ith row of Q, using the following formula: />
Figure FDA0004104019720000021
Figure FDA0004104019720000022
S22, inputting the behavior feature matrix obtained in the step S21, and using a neural collaborative filtering algorithm to represent a learning function of the behavior;
s23, performing independent learning on non-target behaviors and performing multi-behavior learning strategies of target behavior joint learning, and generating a recommendation list for a user according to the result of the joint learning; wherein:
s231, independently learning non-target behaviors;
s2311 user-item is in non-target behavior b j Is represented as a matrix Y j ，Y j The size of (2) is M multiplied by N, M is the number of users, and N is the number of articles;
Figure FDA0004104019720000023
s2312 Each non-target behavior b j All have independent prediction and loss functions, b j The prediction function of (2) is as follows:
Figure FDA0004104019720000024
wherein the method comprises the steps of
Figure FDA0004104019720000025
Is the article in action b j Bias item on->
Figure FDA0004104019720000026
Is a functional representation;
s2313 for Single action b j Likelihood functions of (2) are as follows:
Figure FDA0004104019720000031
wherein the method comprises the steps of
Figure FDA0004104019720000032
Representing a behavior matrix Y j Item 1 in>
Figure FDA0004104019720000033
Representing a behavior matrix Y j An item of 0;
s2314 non-target behavior b j The loss function of (2) is as follows:
Figure FDA0004104019720000034
optimizing the loss function of each non-objective behavioral function independently using a stochastic gradient descent method;
s232, performing joint learning on target behaviors;
s2321. To be usedUser and article in target behavior o w Is represented as a matrix Z w ，Z w The size of (2) is M x N, M is the number of users, N is the number of articles,
Figure FDA0004104019720000035
s2322 for target behavior o w For the sake of this, consider non-target behavior and its relevance;
Figure FDA0004104019720000036
wherein c jw For the confidence coefficient c (b) obtained in step S16 j →o w ) t+1 ，
Figure FDA0004104019720000037
Is the article in behavior o w Bias item on->
Figure FDA0004104019720000038
Is a functional representation;
s2323 for a single target behavior o w Likelihood functions of (2) are as follows:
Figure FDA0004104019720000039
/>
wherein the method comprises the steps of
Figure FDA00041040197200000310
Representing a behavior matrix Z w Item 1 in>
Figure FDA00041040197200000311
Representing a behavior matrix Z w An item of 0;
s2324. the joint probability function of all non-target behaviors is expressed as:
Figure FDA00041040197200000312
s2325, taking the negative logarithm of the probability function, obtaining the following loss function, namely an objective function to be optimized:
Figure FDA00041040197200000313
s.t.
Figure FDA00041040197200000314
wherein lambda is w For target action o w The importance of (2) is obtained through priori knowledge;
s233, constructing a recommendation list according to the prediction results obtained in the S231 and the S232;
predicting the probability of target behavior occurrence according to the previous step
Figure FDA0004104019720000041
Weighting the likelihood of all target behaviors may reflect the preference of user u for item i:
Figure FDA0004104019720000042
and after the R values of all the articles are ranked from large to small, recommending the articles ranked in front to the user.
2. The neural network recommendation method based on multi-user behavior according to claim 1, wherein the output of the neural network in S22 is
Figure FDA0004104019720000043
Wherein the method comprises the steps of
Figure FDA0004104019720000044
For a functional representation with respect to feature vectors, σ is a sigmoid function that converts the output into probabilities.
3. The neural network recommendation method based on multi-user behavior according to claim 2, wherein the following steps are performed
Figure FDA0004104019720000045
Comprising 3 commonly used function representation methods, namely a general matrix decomposition method (GMF), a multi-layer perceptron Method (MLP) and a neural matrix decomposition method (NeuMF), which are described below,
the general matrix factorization method (GMF) is to add weights to each term on the result of matrix factorization:
Figure FDA0004104019720000046
wherein h represents a weight vector;
the multi-layer perceptron Method (MLP) learns by a nonlinear method:
Figure FDA0004104019720000047
…
z L ＝ReLU(W L z L-1 +b L )
Figure FDA0004104019720000048
wherein L represents the number of hidden layers in the multi-layer perceptron, W x And b x The weight and the bias item of the x layer are respectively expressed, and the linear rectification function is used as the activation function by default;
the neural matrix decomposition method (NeuMF) combines GMF and MLP according to a certain weight:
Figure FDA0004104019720000049
where h represents a weight vector, z L Is the result of the last layer of the hidden layer obtained through MLP learning.
4. A neural network recommendation system based on multi-user behaviors, which is applied to the neural network recommendation method based on multi-user behaviors as claimed in any one of claims 1 to 3, and is characterized by comprising a behavior association algorithm and a neural network recommendation model based on multi-user behaviors;
the behavior association algorithm is used for mining the association between the target behavior and obtaining the confidence coefficient from the non-target behavior to the target behavior;
the neural network recommendation model based on the user multi-behavior is used for mining the relation between the multi-behavior and the user preference, wherein a behavior feature matrix of the user and the object is constructed, a neural collaborative filtering model is applied, the model is solved by using a behavior classification learning strategy according to the confidence coefficient obtained by a behavior association algorithm, and finally an object recommendation list is obtained.
CN201911207316.7A 2019-11-29 2019-11-29 Neural network recommendation method based on multi-user behavior Active CN110991601B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911207316.7A CN110991601B (en) 2019-11-29 2019-11-29 Neural network recommendation method based on multi-user behavior

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911207316.7A CN110991601B (en) 2019-11-29 2019-11-29 Neural network recommendation method based on multi-user behavior

Publications (2)

Publication Number Publication Date
CN110991601A CN110991601A (en) 2020-04-10
CN110991601B true CN110991601B (en) 2023-04-21

Family

ID=70088764

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911207316.7A Active CN110991601B (en) 2019-11-29 2019-11-29 Neural network recommendation method based on multi-user behavior

Country Status (1)

Country Link
CN (1) CN110991601B (en)

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111523044B (en) * 2020-07-06 2020-10-23 南京梦饷网络科技有限公司 Method, computing device, and computer storage medium for recommending target objects
CN113763075A (en) * 2020-07-17 2021-12-07 北京沃东天骏信息技术有限公司 Method, device, equipment and medium for pushing articles
CN111859155A (en) * 2020-08-04 2020-10-30 深圳前海微众银行股份有限公司 Item recommendation method, equipment and computer-readable storage medium
CN114630185B (en) * 2020-12-11 2024-05-07 江西省众灿互动科技股份有限公司 Target user identification method and device, electronic equipment and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109446413A (en) * 2018-09-25 2019-03-08 上海交通大学 Serializing recommended method based on item associations relationship
CN109522474A (en) * 2018-10-19 2019-03-26 上海交通大学 Recommended method based on interaction sequence data mining depth user's similitude
CN110188283A (en) * 2019-06-05 2019-08-30 中国人民解放军国防科技大学 Information recommendation method and system based on joint neural network collaborative filtering

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109446413A (en) * 2018-09-25 2019-03-08 上海交通大学 Serializing recommended method based on item associations relationship
CN109522474A (en) * 2018-10-19 2019-03-26 上海交通大学 Recommended method based on interaction sequence data mining depth user's similitude
CN110188283A (en) * 2019-06-05 2019-08-30 中国人民解放军国防科技大学 Information recommendation method and system based on joint neural network collaborative filtering

Also Published As

Publication number Publication date
CN110991601A (en) 2020-04-10

Similar Documents

Publication Publication Date Title
CN110991601B (en) Neural network recommendation method based on multi-user behavior
CN112598462B (en) Personalized recommendation method and system based on collaborative filtering and deep learning
CN109299396B (en) Convolutional neural network collaborative filtering recommendation method and system fusing attention model
Sun et al. What and how: generalized lifelong spectral clustering via dual memory
CN110717098B (en) Meta-path-based context-aware user modeling method and sequence recommendation method
Ortega et al. Recommending items to group of users using matrix factorization based collaborative filtering
Karatzoglou et al. Multiverse recommendation: n-dimensional tensor factorization for context-aware collaborative filtering
Chen et al. Efficient ant colony optimization for image feature selection
CN110781409B (en) Article recommendation method based on collaborative filtering
CN111737578A (en) Recommendation method and system
CN111178986B (en) User-commodity preference prediction method and system
CN114298783A (en) Commodity recommendation method and system based on matrix decomposition and fusion of user social information
Zhou et al. Manifold optimization-assisted Gaussian variational approximation
Gao $ Ae^ 2I $: A Double Autoencoder for Imputation of Missing Values
CN114565436A (en) Vehicle model recommendation system, method, device and storage medium based on time sequence modeling
Ibrahim et al. Recommender systems: algorithms, evaluation and limitations
Zhang et al. High-order factorization machine based on cross weights network for recommendation
CN117252665B (en) Service recommendation method and device, electronic equipment and storage medium
Patil et al. Comparison of different music recommendation system algorithms
El Mouatasim et al. Stochastic perturbation of subgradient algorithm for nonconvex deep neural networks
Monemian A neuroevolutionary neural network-based collaborative filtering recommendation system
Hsu et al. A multistrategy approach to classifier learning from time series
Le MetaRec: Meta-Learning Meets Recommendation Systems
Kitazawa Recommendation. jl: A Framework for Building Recommender Systems in Julia
Deng et al. Real-time collaborative filtering using extreme learning machine

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