CN109767301B - Recommendation method and system, computer device and computer readable storage medium - Google Patents

Recommendation method and system, computer device and computer readable storage medium Download PDF

Info

Publication number
CN109767301B
CN109767301B CN201910032524.1A CN201910032524A CN109767301B CN 109767301 B CN109767301 B CN 109767301B CN 201910032524 A CN201910032524 A CN 201910032524A CN 109767301 B CN109767301 B CN 109767301B
Authority
CN
China
Prior art keywords
target user
interest
user
friend
social network
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
CN201910032524.1A
Other languages
Chinese (zh)
Other versions
CN109767301A (en
Inventor
宋卫平
肖之屏
王一帆
劳伦特·查林
张铭
唐建
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.)
Peking University
Original Assignee
Peking 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 Peking University filed Critical Peking University
Priority to CN201910032524.1A priority Critical patent/CN109767301B/en
Publication of CN109767301A publication Critical patent/CN109767301A/en
Application granted granted Critical
Publication of CN109767301B publication Critical patent/CN109767301B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Abstract

The invention relates to a recommendation method and a recommendation system, wherein the recommendation method comprises the following steps: constructing a social network of the target user corresponding to the consumed item set of the target user; establishing a dynamic personal interest model of a target user according to the item set; constructing a short-term interest model of the social network according to the item set; constructing a long-term interest model of the social network; splicing according to the short-term interest model and the long-term interest model; calculating a node representation of a target user and node representations of friends in a social network; calculating a combined feature weight according to the weight of the friend in the social network about the target user; carrying out nonlinear transformation on the combined characteristic weight; calculating according to the dynamic personal interest model; obtaining the probability of recommending articles according to the final interest of the user; calculating a log-likelihood function value according to the probability of the recommended item; according to the technical scheme, the social relationship of the user and the dynamic interest and hobby factors of the user can be considered at the same time, so that the recommendation accuracy is improved.

Description

Recommendation method and system, computer device and computer readable storage medium
Technical Field
The present invention relates to the field of information recommendation, and in particular, to a recommendation method, a recommendation system, a computer device, and a computer-readable storage medium.
Background
Hidasi et al propose a technique for session-based recommendations using LSTM. The method essentially utilizes the LSTM mentioned above to model the representation of each consumer item from the user's historical consumption records. To recommend the next item, they represent the current interest of the user with the last item they consumed, we calculate how similar the current interest is to all items based on the mathematical representation of this interest, and finally we recommend to the user an item that is closest to the current interest. This technique is a well-established technique, but it has many problems. First it does not model the user's long-term interests, only models his recent interests with the user's recent consumption records. Another problem is that it does not take into account the social impact of the user, which makes the prediction of the model very biased. Chaney et al propose a social cedar decomposition model. They model a trust factor to characterize the recommendation influence degree of a friend on a certain user, and consider the influence of the friend when modeling the user interest, but the model has been proposed for a long time and does not consider the serialization characteristics of the consumption history, so that certain defects exist.
The prior art either models the dynamic interests of the user or analyzes social influences in the recommendation system, but to our knowledge there is no technique to combine these two factors. One recent study is about simulating user behavior at the session level using a recurrent neural network, but does not consider social effects. Other work has investigated social effects, e.g., Ma et al explore the systematic effects of social relationships of friends on recommendations. However, the contributions from different users are static and they do not change according to the current interests of the recommended user.
Disclosure of Invention
The present invention is directed to solving at least one of the problems of the prior art or the related art.
Therefore, an object of the present invention is to provide a recommendation method based on session and social influence, which can simultaneously consider the social relationship of the user and the dynamic interest and hobby factors of the user to improve the accuracy of recommendation; and according to the interests and hobbies of the user, the friends with higher influence in the social relationship are dynamically deduced, so that the friends with similar interests have larger influence on the recommendation result.
Another object of the present invention is to provide a recommendation system based on conversation and social influence, which can comprehensively consider the interests of the user and the interests of friends thereof when generating a recommendation result, and add a dynamic inference module for influence on the friends, so as to better capture the recent interests of the related friends and serve the object recommendation of the target user.
It is yet another object of the present invention to provide a computer apparatus.
It is yet another object of the present invention to provide a computer-readable storage medium.
In order to achieve the above object, a technical solution of a first aspect of the present invention provides a recommendation method based on session and social influence, including the following steps:
building a set of items consumed in a current session of a target user
Figure GDA0002946692340000021
And a social network G corresponding to the target user, such that:
Figure GDA0002946692340000022
G＝(U，E)；
wherein U represents a target user, U represents a friend set of the target user U, E represents a social relationship with the target user U, and i represents a consumed commodity;
according to the collection of articles
Figure GDA0002946692340000023
Establishing a dynamic personal interest model of a target user, and enabling:
Figure GDA0002946692340000024
wherein h isnRepresents the latest interest of the target user, hn-1The previous interest representing the latest interest, f represents the latest consumed commodity of the target user u
Figure GDA0002946692340000031
A non-linear function combined with a previous interest of the latest interest;
according to the collection of articles
Figure GDA0002946692340000032
Building short-term interest model of social network G
Figure GDA0002946692340000033
Order:
Figure GDA0002946692340000034
wherein the content of the first and second substances,
Figure GDA0002946692340000035
representing the latest interests of friend k of target user u,
Figure GDA0002946692340000036
representing the previous interest of the friend k,
Figure GDA0002946692340000037
the latest consumed item representing friend k of target user u;
building a long-term interest model for social network G
Figure GDA0002946692340000038
Order:
Figure GDA0002946692340000039
wherein, the long-term interest model of friend k in social network G
Figure GDA00029466923400000310
Is the embedded representation matrix W of the target user uuThe kth line of (1);
according to short-term interest model
Figure GDA00029466923400000311
And long-term interest model
Figure GDA00029466923400000312
Splicing to obtain a spliced model skLet us order
Figure GDA00029466923400000313
Wherein relu (x) max (0, x) is a non-linear activation function, W1Is a transformation matrix;
computing node representations of target users
Figure GDA00029466923400000314
And node representation of friend k in social network G
Figure GDA00029466923400000315
Order:
Figure GDA00029466923400000316
wherein the content of the first and second substances,
Figure GDA00029466923400000317
and
Figure GDA00029466923400000318
respectively as a representation of the target user u and his friend k at level l,
Figure GDA00029466923400000319
to calculate the function of the similarity of two nodes,
Figure GDA00029466923400000320
the weight value of the user k in the social network G about the target user u is obtained;
calculating a combined feature weight according to the weight of the friend k in the social network G about the target user u, and ordering:
Figure GDA0002946692340000041
wherein the content of the first and second substances,
Figure GDA0002946692340000042
is the fusion of the interests of the social network G of the target user u on level i;
and carrying out nonlinear transformation on the combined characteristic weight to obtain:
Figure GDA0002946692340000043
wherein, W(l)Is a shared and learnable weight matrix of the l layers;
calculating to obtain the final interest of the user according to the dynamic personal interest model, and enabling:
Figure GDA0002946692340000044
wherein, W2In the form of a linear transformation matrix, the transformation matrix,
Figure GDA0002946692340000045
is the final interest of the target user;
the probability that the recommended item is y is obtained according to the final interest of the user, namely:
Figure GDA0002946692340000046
wherein N (u) is the number of users in social network G, zyFor the embedded representation of the item y,Ithe number of the articles is the number of the articles,
Figure GDA0002946692340000047
a set of consumed items for each friend k;
calculating a log-likelihood function value for the item according to the probability that the recommended item is y:
Figure GDA0002946692340000048
in the technical scheme, the method aims at simultaneously modeling the dynamic interest and the real-time social influence of the user. Specifically, a neural network technology is used for modeling a current conversation of a user, the interest preference of the user is extracted, the influence of friends of the user on the user in a current scene is calculated in real time, and article recommendation is carried out according to the own preference and the influence of the friends.
In the above technical solution, preferably, the non-linear function f combining the latest consumed commodity with the previous interest is:
Figure GDA0002946692340000049
Figure GDA00029466923400000410
Figure GDA0002946692340000051
Figure GDA0002946692340000052
Figure GDA0002946692340000053
Figure GDA0002946692340000054
wherein, sigma is sigmoid function, sigma (x) ═ 1+ exp (-x)-1，{Wx，Wf，Wo，WcExpressed as a learnable parameter matrix in a recurrent neural network, { bx，bf，bo，bcIs learnable in a recurrent neural networkA parameter vector; { xn，fn，onThe input gate, the forgetting gate and the output gate in the recurrent neural network are respectively; c. Cn' represents an intermediate variable transformed by a neural network, cnAnd cn-1Respectively representing the current time value and the last time value of a cell memory in the recurrent neural network; hnRepresents the passing of the value in the memory of the cell through the output gate onAfter the action of (3), the output vector of the neural network at the current moment is circulated.
The technical scheme of the second aspect of the invention provides a recommendation system based on conversation and social influence, which comprises the following steps:
a construction module arranged to construct a set of items consumed in a current session of a target user
Figure GDA0002946692340000055
And a social network G corresponding to the target user, such that:
Figure GDA0002946692340000056
G＝(U，E)；
wherein U represents a target user, U represents a friend set of the target user U, E represents a social relationship with the target user U, and i represents a consumed commodity;
a dynamic personal interest model building module arranged for aggregating according to the items
Figure GDA0002946692340000057
Establishing a dynamic personal interest model of a target user, and enabling:
Figure GDA0002946692340000058
wherein h isnRepresents the latest interest of the target user, hn-1A previous interest representing the latest interest, f represents a non-linear function of the target user u combining the latest consumed commodity with the previous interest of the latest interest;
a short term interest model building module configured to build a set of items from the set of items
Figure GDA0002946692340000061
Building short-term interest model of social network G
Figure GDA0002946692340000062
Order:
Figure GDA0002946692340000063
wherein the content of the first and second substances,
Figure GDA0002946692340000064
representing the latest interests of friend k of target user u,
Figure GDA0002946692340000065
representing the previous interest of the friend k,
Figure GDA0002946692340000066
the latest consumed item representing friend k of target user u;
a long-term interest model construction module configured to construct a long-term interest model of the social network G
Figure GDA0002946692340000067
Order:
Figure GDA0002946692340000068
wherein, the long-term interest model of friend k in social network G
Figure GDA0002946692340000069
Is the embedded representation matrix W of the target user uuThe kth line of (1);
a stitching module configured to generate a short-term interest model based on the short-term interest model
Figure GDA00029466923400000610
And long-term interest model
Figure GDA00029466923400000611
Splicing to obtain a spliced model skLet us order
Figure GDA00029466923400000612
Wherein relu (x) max (0, x) is a non-linear activation function, W1Is a transformation matrix;
a calculation module arranged for calculating a node representation of a target user
Figure GDA00029466923400000613
And node representation of friend k in social network G
Figure GDA00029466923400000614
Order:
Figure GDA00029466923400000615
wherein the content of the first and second substances,
Figure GDA00029466923400000616
and
Figure GDA00029466923400000617
respectively as a representation of the target user u and his friend k at level l,
Figure GDA00029466923400000618
to calculate the function of the similarity of two nodes,
Figure GDA00029466923400000619
the weight value of the user k in the social network G about the target user u is obtained;
a merged feature weight calculation module configured to calculate a merged feature weight according to a weight of a friend k in the social network G with respect to the target user u, such that:
Figure GDA00029466923400000620
wherein the content of the first and second substances,
Figure GDA0002946692340000071
is the fusion of the interests of the social network G of the target user u on level i;
a nonlinear transformation module configured to perform nonlinear transformation on the combined feature weight to obtain:
Figure GDA0002946692340000072
wherein, W(l)Is a shared and learnable weight matrix of the l layers;
a final interest calculation module configured to calculate a final interest of the user according to the dynamic personal interest model, such that:
Figure GDA0002946692340000073
wherein, W2In the form of a linear transformation matrix, the transformation matrix,
Figure GDA0002946692340000074
is the final interest of the target user;
a probability calculation module configured to derive a probability that the recommended item is y according to the final interest of the user, namely:
Figure GDA0002946692340000075
wherein N (u) is the number of users in social network G, zyFor the embedded representation of the item y,Ithe number of the articles is the number of the articles,
Figure GDA0002946692340000076
a set of consumed items for each friend k;
a log-likelihood function value calculation module arranged to calculate a log-likelihood function value for the item according to the probability that the recommended item is y:
Figure GDA0002946692340000077
in the technical scheme, in the recommendation system, the social relationship of the user and the dynamic interest and hobby factors of the user are considered at the same time, so that the recommendation accuracy is improved; and according to the interests and hobbies of the user, the friends with higher influence in the social relationship are dynamically deduced, so that the friends with similar interests have larger influence on the recommendation result. In order to improve the accuracy of an online platform recommendation system, the invention provides modeling on dynamic interest and hobbies and dynamic social relations of a user. When a recommendation result is generated, the interests of the user and the interests of friends of the user are comprehensively considered, and a dynamic inference module for influencing the friends is added, so that the recent hobbies of the related friends can be better captured, and the object recommendation service is provided for the object recommendation of the target user.
In the above technical solution, preferably, the non-linear function f combining the latest consumed commodity with the previous interest is:
Figure GDA0002946692340000081
Figure GDA0002946692340000082
Figure GDA0002946692340000083
Figure GDA0002946692340000084
Figure GDA0002946692340000085
Figure GDA0002946692340000086
wherein, sigma is sigmoid function, sigma (x) ═ 1+ exp (-x)-1，{Wx，Wf，Wo，WcExpressed as a learnable parameter matrix in a recurrent neural network, { bx，bf，bo，bcThe parameter vector can be learned in the recurrent neural network; { xn，fn，onThe input gate, the forgetting gate and the output gate in the recurrent neural network are respectively; c. Cn' represents an intermediate variable transformed by a neural network, cnAnd cn-1Respectively representing the current time value and the last time value of a cell memory in the recurrent neural network; hnRepresents the passing of the value in the memory of the cell through the output gate onAfter the action of (3), the output vector of the neural network at the current moment is circulated.
An aspect of the third aspect of the present invention provides a computer apparatus, including a processor, configured to implement the steps of the recommendation method of any one of the aspects presented in the first aspect of the present invention when the processor executes a computer program stored in a memory.
An aspect of the fourth aspect of the present invention provides a computer-readable storage medium on which a computer program (instructions) is stored, the computer program (instructions) implementing the steps of the recommendation method of any one of the aspects set forth in the first aspect of the present invention when executed by a processor.
Drawings
The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 illustrates a block flow diagram of a recommendation method in accordance with an embodiment of the present invention;
FIG. 2 shows a block flow diagram of a recommendation system according to another embodiment of the present invention;
FIG. 3 illustrates a dynamic graph attention model diagram in accordance with an embodiment of the present invention;
fig. 4 shows a dynamic graph attention network diagram in accordance with an embodiment of the present invention.
Detailed Description
In order that the above objects, features and advantages of the present invention can be more clearly understood, a more particular description of the invention will be rendered by reference to the appended drawings. It should be noted that the embodiments and features of the embodiments of the present application may be combined with each other without conflict.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced in other ways than those specifically described herein, and therefore the scope of the present invention is not limited to the specific embodiments disclosed below.
Recommendation methods and systems, computer devices, computer-readable storage media according to some embodiments of the invention are described below with reference to fig. 1-4.
To be able to provide effective suggestions to users of an online community, we propose to model both the dynamic interests and the context-dependent social impact of the users. We define the final problem as follows:
the definition (session level social recommendation) has U representing the user set, i representing the consumed goods, and G ═ G, E is the social network, where E is the social relationship with the target user U. For user u, a new set of items to be consumed is given
Figure GDA0002946692340000091
The goal of session-level social recommendations is to simultaneously exploit the dynamic interests of user u (from
Figure GDA0002946692340000092
Information of) and social influence (from
Figure GDA0002946692340000101
Where n (u) is the number of users in social network G) to recommend a subset of I, the items in the subset being items that user u may be interested in at step n + 1.
To this end, the present invention proposes a novel Dynamic Graph attention model (DGRec), as shown in FIG. 3, which can simultaneously model the user's own preferences and the preferences of the user's friends.
DGRec consists of four modules. The first module is a Recurrent Neural Network (RNN) that can model the sequence of items consumed in the user's current session. The interests of the friends of the user are modeled by a combination of their short-term preferences and long-term preferences. Short-term preferences, such as items in their recent sessions, also use RNN coding. The long-term preferences of friends are encoded by learned personal embedded representations. The model then uses a graph attention network to combine the representation of the current user with the representation of his friends. This is a key part of our model and contribution: our proposed mechanism learns to measure the impact of each friend based on the user's current interests. In the last step, the model generates recommendation results by combining the user's current preferences with the (context-dependent) social impact he is subjected to.
As shown in fig. 1, a recommendation method based on conversation and social influence according to an embodiment of the present invention includes the following steps:
s100, constructing a consumed item set in the current conversation of the target user
Figure GDA0002946692340000102
And a social network G corresponding to the target user, such that:
Figure GDA0002946692340000103
G＝(U，E)；
wherein U represents a target user, U represents a friend set of the target user U, E represents a social relationship with the target user U, and i represents a consumed commodity;
to capture the user's rapidly changing interests, we use the RNN to model the actions (i.e., clicks) of the (target) user in the current session. RNNs are standard sequence modeling tools and have recently been used to model user (sequence) preference data. The RNN can infer a set of consumed items input by input
Figure GDA0002946692340000111
Is shown. It can recursively combine all previous and latest input representations, namely: s200, according to the item set
Figure GDA0002946692340000112
Establishing a dynamic personal interest model of the target user
Figure GDA0002946692340000113
Wherein h isnRepresents the latest interest of the target user, hn-1A previous interest representing the latest interest, f represents a non-linear function of the target user u combining the latest consumed commodity with the previous interest of the latest interest;
we believe that in social network G, the user may be affected by recent interests of friends. For this reason, we model the short-term and long-term interests of friends in different ways.
We use the sequence of items recently consumed by a friend (e.g., the friend's latest online conversation) to model his short-term interests. The long-term interest represents the overall interest of a friend and is modeled using an embedded representation of the individual.
For a current conversation of a target user
Figure GDA0002946692340000114
His friends' short-term interests in their respective meetingsThe session before session T + 1. Actions of each friend k
Figure GDA0002946692340000115
Modeling was performed using RNN. In fact, here we re-use the RNN modeling the target user session to model the friend's session. In other words, both RNNs share the same parameters. We denote the short-term preferences of friend k with the final output of RNN, S300, from the item set
Figure GDA0002946692340000116
Building short-term interest model of social network G
Figure GDA0002946692340000117
Order:
Figure GDA0002946692340000118
wherein the content of the first and second substances,
Figure GDA0002946692340000119
representing the latest interests of friend k of target user u,
Figure GDA00029466923400001110
representing the previous interest of the friend k,
Figure GDA00029466923400001111
the latest consumed item representing friend k of target user u;
the long-term preferences of friends reflect their overall interests. Since long-term preferences are not time sensitive, we use a vector to represent them.
S400, constructing a long-term interest model of the social network G
Figure GDA00029466923400001112
Order:
Figure GDA00029466923400001113
wherein, the long-term interest model of friend k in social network G
Figure GDA0002946692340000121
Is the embedded representation matrix W of the target user uuThe kth line of (1);
s500, according to the short-term interest model
Figure GDA0002946692340000122
And long-term interest model
Figure GDA0002946692340000123
Splicing to obtain a spliced model skLet us order
Figure GDA0002946692340000124
Wherein relu (x) max (0, x) is a non-linear activation function, W1Is a transformation matrix;
we use a novel attention network to obtain a mixed representation of the interests of the target user and the interests of his friends. First, we encode a friendship network in the graph, where the nodes correspond to users (i.e., the target user and their friends) and the edges represent friendships. Furthermore, each node uses its representation of the corresponding user as a (dynamic) feature. Second, these features are propagated edge-wise using a message passing algorithm. The main novelty of our approach is to use an attention mechanism to weigh the features that propagate along each edge. The weight corresponds to the degree of influence of the friend on the target user. After a fixed number of message delivery iterations, the resulting characteristic at the target user node is the combined representation.
For each user, we construct a graph with the user and his friends as nodes. If user u has | N (u) | friends, then the graph has | N (u) | +1 nodes. Initial representation h of user unIs taken as an initial feature of node u
Figure GDA0002946692340000125
(each time user u consumes a new presence
Figure GDA0002946692340000126
The feature is updated after the item). For friend k, the corresponding node feature is set to SkAnd remains unchanged for the time stamp T + 1. Formally, the characteristic representation of a node is
Figure GDA0002946692340000127
We propose a novel dynamic graph attention network that models context-dependent social influence and uses an attention mechanism to guide the propagation of the influence. The entire process is illustrated in fig. 4.
S600, calculating the node representation of the target user
Figure GDA0002946692340000128
And node representation of friend k in social network G
Figure GDA0002946692340000129
Order:
Figure GDA0002946692340000131
wherein the content of the first and second substances,
Figure GDA0002946692340000132
and
Figure GDA0002946692340000133
respectively as a representation of the target user u and his friend k at level l,
Figure GDA0002946692340000134
to calculate the function of the similarity of two nodes,
Figure GDA0002946692340000135
weight of user k in social network G with respect to target user u (as existing)
Figure GDA0002946692340000136
Background), or the level of influence;
s700, calculating a combined feature weight according to the weight of the friend k in the social network G about the target user u, and ordering:
Figure GDA0002946692340000137
wherein the content of the first and second substances,
Figure GDA0002946692340000138
is the fusion of the interests of the social network G of the target user u on level i; s800, carrying out nonlinear transformation on the combined feature weight to obtain:
Figure GDA0002946692340000139
wherein, W(l)Is a shared and learnable weight matrix of l layers, where each layer represents a convolution operation of the graph convolution network, and the value gets the final representation of each node after l layers of convolution from the first layer. Representation of Merge (social impact) We use
Figure GDA00029466923400001310
To indicate.
Since the interest of a user is determined by both his recent behavior and social influence, his final representation is obtained by merging both of the fully connected layers, i.e. S900, the final interest of the user is calculated according to a dynamic personal interest model, let:
Figure GDA00029466923400001311
wherein, W2In the form of a linear transformation matrix, the transformation matrix,
Figure GDA00029466923400001312
is the final interest of the target user;
s1000, obtaining the probability that the recommended item is y according to the final interest of the user, namely:
Figure GDA00029466923400001313
wherein N (u) is the number of users in social network G, zyFor embedded representation of item y, | I | is the number of items,
Figure GDA0002946692340000141
a set of consumed items for each friend k;
s1100, calculating a log-likelihood function value of the item according to the probability that the recommended item is y:
Figure GDA0002946692340000142
the function is optimized using a gradient descent method.
In this embodiment, the method aims to model both the dynamic interests and the real-time social impact of the user. Specifically, a neural network technology is used for modeling a current conversation of a user, the interest preference of the user is extracted, the influence of friends of the user on the user in a current scene is calculated in real time, and article recommendation is carried out according to the own preference and the influence of the friends.
As shown in FIG. 2, a recommendation system 1000 based on conversational and social influence according to another embodiment of the invention includes:
a construction module 10 arranged for constructing a set of items consumed in a current session of a target user
Figure GDA0002946692340000143
And a social network G corresponding to the target user, such that:
Figure GDA0002946692340000144
G＝(U，E)；
wherein U represents a target user, U represents a friend set of the target user U, E represents a social relationship with the target user U, and i represents a consumed commodity;
a dynamic personal interest model building module 20 arranged for aggregating according to the items
Figure GDA0002946692340000145
Establishing a dynamic personal interest model of a target user, and enabling:
Figure GDA0002946692340000146
wherein h isnRepresents the latest interest of the target user, hn-1The previous interest representing the latest interest, f represents the latest consumed commodity of the target user u
Figure GDA0002946692340000147
A non-linear function combined with a previous interest of the latest interest;
a short term interest model building module 30 arranged for assembling from the items
Figure GDA0002946692340000148
Building short-term interest model of social network G
Figure GDA0002946692340000149
Order:
Figure GDA0002946692340000151
wherein the content of the first and second substances,
Figure GDA0002946692340000152
representing the latest interests of friend k of target user u in social network G,
Figure GDA0002946692340000153
representing the previous interest of the friend k,
Figure GDA0002946692340000154
the latest consumed item representing friend k of target user u;
a long-term interest model construction module 40 arranged for constructing a long-term interest model of the social network G
Figure GDA0002946692340000155
Order:
Figure GDA0002946692340000156
wherein, the long-term interest model of friend k in social network G
Figure GDA0002946692340000157
Is the embedded representation matrix W of the target user uuThe kth line of (1);
a stitching module 50 arranged for generating a short-term interest model
Figure GDA0002946692340000158
And long-term interest model
Figure GDA0002946692340000159
Splicing to obtain a spliced model skLet us order
Figure GDA00029466923400001510
Wherein relu (x) max (0, x) is a non-linear activation function, W1Is a transformation matrix;
a calculation module 60 arranged for calculating a node representation of a target user
Figure GDA00029466923400001511
And node representation of friend k in social network G
Figure GDA00029466923400001512
Order:
Figure GDA00029466923400001513
wherein the content of the first and second substances,
Figure GDA00029466923400001514
and
Figure GDA00029466923400001515
respectively as a representation of the target user u and his friend k at level l,
Figure GDA00029466923400001516
to calculate the function of the similarity of two nodes,
Figure GDA00029466923400001517
the weight value of the user k in the social network G about the target user u is obtained;
a merged feature weight calculation module 70 configured to calculate a merged feature weight according to the weight of friend k in social network G with respect to target user u, such that:
Figure GDA00029466923400001518
wherein the content of the first and second substances,
Figure GDA00029466923400001519
is the fusion of the interests of the social network G of the target user u on level i;
a nonlinear transformation module 80 configured to perform nonlinear transformation on the combined feature weight to obtain:
Figure GDA0002946692340000161
wherein, W(l)Is a shared and learnable weight matrix of the l layers;
a final interest calculation module 90 configured to calculate a final interest of the user according to the dynamic personal interest model, such that:
Figure GDA0002946692340000162
wherein, W2In the form of a linear transformation matrix, the transformation matrix,
Figure GDA0002946692340000163
is the final interest of the target user;
a probability calculation module 100 configured to derive a probability that the recommended item is y according to the final interest of the user, namely:
Figure GDA0002946692340000164
wherein N (u) is the number of users in social network G, zyFor embedded representation of item y, | I | is the number of items,
Figure GDA0002946692340000165
a set of consumed items for each friend k;
a log-likelihood function value calculation module 110 arranged to calculate a log-likelihood function value for an item according to the probability that the recommended item is y:
Figure GDA0002946692340000166
in the embodiment, in the recommendation system, the social relationship of the user and the dynamic interest and hobby factors of the user are considered at the same time, so that the recommendation accuracy is improved; and according to the interests and hobbies of the user, the friends with higher influence in the social relationship are dynamically deduced, so that the friends with similar interests have larger influence on the recommendation result. In order to improve the accuracy of an online platform recommendation system, the invention provides modeling on dynamic interest and hobbies and dynamic social relations of a user. When a recommendation result is generated, the interests of the user and the interests of friends of the user are comprehensively considered, and a dynamic inference module for influencing the friends is added, so that the recent hobbies of the related friends can be better captured, and the object recommendation service is provided for the object recommendation of the target user.
In any of the above embodiments, preferably, the non-linear function f that combines the latest consumed good with the previous interest is:
Figure GDA0002946692340000171
Figure GDA0002946692340000172
Figure GDA0002946692340000173
Figure GDA0002946692340000174
Figure GDA0002946692340000175
Figure GDA0002946692340000176
wherein, sigma is sigmoid function, sigma (x) ═ 1+ exp (-x)-1，{Wx，Wf，Wo，WcExpressed as a learnable parameter matrix in a recurrent neural network, { bx，bf，bo，bcThe parameter vector can be learned in the recurrent neural network; { xn，fn，onThe input gate, the forgetting gate and the output gate in the recurrent neural network are respectively; c. Cn' represents an intermediate variable transformed by a neural network, cnAnd cn-1Respectively representing the current time value and the last time value of a cell memory in the recurrent neural network; hnRepresents the passing of the value in the memory of the cell through the output gate onAfter the action of (3), the output vector of the neural network at the current moment is circulated.
The computer apparatus according to a further embodiment of the present invention comprises a processor for implementing the steps of the recommendation method according to any one of the aspects set forth in the first aspect of the present invention when executing the computer program stored in the memory.
The computer-readable storage medium of the further embodiment of the present invention stores thereon a computer program (instructions) which, when executed by a processor, implements the steps of the recommendation method of any one of the aspects set forth in the first aspect of the present invention.
In the present invention, the terms "mounting," "connecting," "fixing," and the like are used in a broad sense, for example, "connecting" may be a fixed connection, a detachable connection, or an integral connection; "coupled" may be direct or indirect through an intermediary. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations.
In the description of the present invention, it is to be understood that the terms "upper", "lower", "left", "right", "front", "rear", and the like indicate orientations or positional relationships based on those shown in the drawings, and are only for convenience of description and simplification of description, but do not indicate or imply that the referred device or unit must have a specific direction, be constructed in a specific orientation, and be operated, and thus, should not be construed as limiting the present invention.
In the description herein, the description of the terms "one embodiment," "some embodiments," "specific embodiments," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. 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 recommendation method based on conversation and social influence is characterized by comprising the following steps:
building a set of items consumed in a current session of a target user
Figure FDA0002946692330000011
And a social network G corresponding to the target user, and order:
Figure FDA0002946692330000012
G＝(U，E)；
wherein U represents the target user, U represents a set of friends of the target user U, E represents a social relationship with the target user U, and i represents a consumed commodity;
according to the item set
Figure FDA0002946692330000013
Establishing a dynamic personal interest model of the target user, and enabling:
Figure FDA0002946692330000014
wherein h isnRepresenting the latest interest of the target user, hn-1The previous interest representing the latest interest, f representing that the target user u will be the latest consumed commodity
Figure FDA0002946692330000015
A non-linear function combined with a previous one of the latest interests;
according to the item set
Figure FDA0002946692330000016
Constructing a short-term interest model of the social network G
Figure FDA0002946692330000017
Order:
Figure FDA0002946692330000018
wherein the content of the first and second substances,
Figure FDA0002946692330000019
representing the latest interests of friend k of the target user u,
Figure FDA00029466923300000110
representing the previous interest of the friend k,
Figure FDA00029466923300000111
the latest consumed item representing friend k of the target user u;
building a long-term interest model of the social network G
Figure FDA00029466923300000112
Order:
Figure FDA00029466923300000113
wherein the long-term interest model of friend k in the social network G
Figure FDA00029466923300000114
Is the embedded representation matrix W of the target user uuThe kth line of (1);
according to the short-term interest model
Figure FDA00029466923300000115
And the long-term interest model
Figure FDA00029466923300000116
Splicing to obtain a spliced model skLet us order
Figure FDA0002946692330000021
Wherein relu (x) max (0, x) is a non-linear activation function, W1Is a transformation matrix;
computing a node representation of the target user
Figure FDA0002946692330000022
And a node representation of friend k in the social network G
Figure FDA0002946692330000023
Order:
Figure FDA0002946692330000024
wherein the content of the first and second substances,
Figure FDA0002946692330000025
and
Figure FDA0002946692330000026
respectively as a representation of the target user u and his friend k at level l,
Figure FDA0002946692330000027
to calculate the function of the similarity of two nodes,
Figure FDA0002946692330000028
the weight value of the user k in the social network G about the target user u is obtained;
calculating a combined feature weight according to the weight of the friend k in the social network G about the target user u, and ordering:
Figure FDA0002946692330000029
wherein the content of the first and second substances,
Figure FDA00029466923300000210
is a fusion of the interests of the social network G of the target user u on level l;
and carrying out nonlinear transformation on the combined feature weight to obtain:
Figure FDA00029466923300000211
wherein, W(l)Is a shared and learnable weight matrix of the l layers;
calculating to obtain the final interest of the user according to the dynamic personal interest model, and ordering:
Figure FDA00029466923300000212
wherein, W2In the form of a linear transformation matrix, the transformation matrix,
Figure FDA00029466923300000213
is the final interest of the target user;
obtaining the probability that the recommended item is y according to the final interest of the user, namely:
Figure FDA00029466923300000214
wherein N (u) is the number of users in the social network G, zyFor embedded representation of item y, | I | is the number of items,
Figure FDA00029466923300000215
a set of consumed items for each friend k;
calculating a log-likelihood function value for the item according to the probability that the recommended item is y:
Figure FDA00029466923300000216
2. the method of claim 1, wherein the method comprises: the non-linear function f that combines the latest consumed good with the previous interest is:
Figure FDA0002946692330000031
Figure FDA0002946692330000032
Figure FDA0002946692330000033
Figure FDA0002946692330000034
Figure FDA0002946692330000035
Figure FDA0002946692330000036
wherein, sigma is sigmoid function, sigma (x) ═ 1+ exp (-x)-1，{Wx，Wf，Wo，WcExpressed as a learnable parameter matrix in a recurrent neural network, { bx，bf，bo，bcThe parameter vector can be learned in the recurrent neural network; { xn，fn，onThe input gate, the forgetting gate and the output gate in the recurrent neural network are respectively; c. Cn' represents an intermediate variable transformed by a neural network, cnAnd cn-1Respectively representing the current time value and the last time value of a cell memory in the recurrent neural network; hnRepresents the passing of the value in the memory of the cell through the output gate onAfter the action of (3), the output vector of the neural network at the current moment is circulated.
3. A recommendation system based on conversational and social influence, comprising:
a construction module arranged to construct a set of items consumed in a current session of a target user
Figure FDA0002946692330000037
And a social network G corresponding to the target user, and order:
Figure FDA0002946692330000038
G＝(U，E)；
wherein U represents the target user, U represents a set of friends of the target user U, E represents a social relationship with the target user U, and i represents a consumed commodity;
a dynamic personal interest model building module configured to build a set of items from the set of items
Figure FDA0002946692330000039
Establishing a dynamic personal interest model of the target user, and enabling:
Figure FDA00029466923300000310
wherein h isnRepresenting the latest interest of the target user, hn-1The previous interest representing the latest interest, f representing that the target user u will be the latest consumed commodity
Figure FDA0002946692330000041
A non-linear function combined with a previous one of the latest interests;
a short term interest model building module configured to build a short term interest model based on the set of items
Figure FDA0002946692330000042
Constructing a short-term interest model of the social network G
Figure FDA0002946692330000043
Order:
Figure FDA0002946692330000044
wherein the content of the first and second substances,
Figure FDA0002946692330000045
representing the latest interests of friend k of the target user u,
Figure FDA0002946692330000046
representing the previous interest of the friend k,
Figure FDA0002946692330000047
the latest consumed item representing friend k of the target user u;
a long-term interest model construction module arranged for constructing a long-term interest model of the social network G
Figure FDA0002946692330000048
Order:
Figure FDA0002946692330000049
wherein the long-term interest model of friend k in the social network G
Figure FDA00029466923300000410
Is the embedded representation matrix W of the target user uuThe kth line of (1);
a stitching module configured to apply a short-term interest model to the data stream
Figure FDA00029466923300000411
And the long-term interest model
Figure FDA00029466923300000412
Splicing to obtain a spliced model skLet us order
Figure FDA00029466923300000413
Wherein relu (x) max (0, x) is a non-linear activation function, W1Is a transformation matrix;
a calculation module arranged for calculating a node representation of the target user
Figure FDA00029466923300000414
And a node representation of friend k in the social network G
Figure FDA00029466923300000415
Order:
Figure FDA00029466923300000416
wherein the content of the first and second substances,
Figure FDA00029466923300000417
and
Figure FDA00029466923300000418
respectively as a representation of the target user u and his friend k at level l,
Figure FDA00029466923300000419
to calculate the function of the similarity of two nodes,
Figure FDA00029466923300000420
the weight value of the user k in the social network G about the target user u is obtained;
a merged feature weight calculation module configured to calculate a merged feature weight according to the weight of friend k in the social network G with respect to the target user u, such that:
Figure FDA0002946692330000051
wherein the content of the first and second substances,
Figure FDA0002946692330000052
is a fusion of the interests of the social network G of the target user u on level l;
a nonlinear transformation module configured to perform nonlinear transformation on the merged feature weight to obtain:
Figure FDA0002946692330000053
wherein, W(l)Is a shared and learnable weight matrix of the l layers;
a final interest calculating module configured to calculate a final interest of the user according to the dynamic personal interest model, such that:
Figure FDA0002946692330000054
wherein, W2In the form of a linear transformation matrix, the transformation matrix,
Figure FDA0002946692330000055
is the final interest of the target user;
a probability calculation module configured to derive a probability that a recommended item is y based on the user's ultimate interest, namely:
Figure FDA0002946692330000056
wherein N (u) is the number of users in the social network G, zyFor embedded representation of item y, | I | is the number of items,
Figure FDA0002946692330000057
a set of consumed items for each friend k;
a log-likelihood function value calculation module arranged to calculate a log-likelihood function value for the item according to the probability that the recommended item is y:
Figure FDA0002946692330000058
4. the conversational and social influence based recommendation system according to claim 3, wherein: the non-linear function f that combines the latest consumed good with the previous interest is:
Figure FDA0002946692330000059
Figure FDA00029466923300000510
Figure FDA00029466923300000511
Figure FDA00029466923300000512
Figure FDA0002946692330000061
Figure FDA0002946692330000062
wherein, sigma is sigmoid function, sigma (x) ═ 1+ exp (-x)-1，{Wx，Wf，Wo，WcExpressed as a learnable parameter matrix in a recurrent neural network, { bx，bf，bo，bcThe parameter vector can be learned in the recurrent neural network; { xn，fn，onThe input gate, the forgetting gate and the output gate in the recurrent neural network are respectively; c. Cn' represents intermediate variable transformed by neural network，cnAnd cn-1Respectively representing the current time value and the last time value of a cell memory in the recurrent neural network; hnRepresents the passing of the value in the memory of the cell through the output gate onAfter the action of (3), the output vector of the neural network at the current moment is circulated.
5. A computer arrangement comprising a processor for implementing the steps of the method as claimed in claim 1 or 2 when executing a computer program stored in a memory.
6. A computer-readable storage medium having stored thereon a computer program, characterized in that: the computer program realizes the steps of the recommendation method as claimed in claim 1 or 2 when executed by a processor.
CN201910032524.1A 2019-01-14 2019-01-14 Recommendation method and system, computer device and computer readable storage medium Active CN109767301B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910032524.1A CN109767301B (en) 2019-01-14 2019-01-14 Recommendation method and system, computer device and computer readable storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910032524.1A CN109767301B (en) 2019-01-14 2019-01-14 Recommendation method and system, computer device and computer readable storage medium

Publications (2)

Publication Number Publication Date
CN109767301A CN109767301A (en) 2019-05-17
CN109767301B true CN109767301B (en) 2021-05-07

Family

ID=66453656

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910032524.1A Active CN109767301B (en) 2019-01-14 2019-01-14 Recommendation method and system, computer device and computer readable storage medium

Country Status (1)

Country Link
CN (1) CN109767301B (en)

Families Citing this family (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110147911B (en) * 2019-05-28 2022-09-09 吉林大学 Social influence prediction model and prediction method based on content perception
CN110245293B (en) * 2019-05-28 2023-07-11 深圳市雅阅科技有限公司 Network content recall method and device
CN110502701B (en) * 2019-07-08 2022-03-25 华南师范大学 Friend recommendation method, system and storage medium introducing attention mechanism
CN111125445B (en) * 2019-12-17 2023-08-15 北京百度网讯科技有限公司 Community theme generation method and device, electronic equipment and storage medium
CN111125537B (en) * 2019-12-31 2020-12-22 中国计量大学 Session recommendation method based on graph representation
CN111222054B (en) * 2020-01-03 2020-12-11 中国计量大学 Session social contact recommendation method based on context neighbor relation modeling
CN111258593B (en) * 2020-01-08 2023-10-27 恒安嘉新（北京）科技股份公司 Application program prediction model building method and device, storage medium and terminal
CN111260447A (en) * 2020-02-11 2020-06-09 深圳前海达闼云端智能科技有限公司 Commodity recommendation method and device for vending robot
CN111460331B (en) * 2020-04-07 2020-12-25 中国计量大学 Conversation recommendation method based on space-time diagram
CN112163147A (en) * 2020-06-09 2021-01-01 中森云链(成都)科技有限责任公司 Recommendation method for website session scene
CN111782765A (en) * 2020-06-24 2020-10-16 安徽农业大学 Recommendation method based on graph attention machine mechanism
CN112445981A (en) * 2020-11-04 2021-03-05 西安电子科技大学 Social and consumption joint recommendation system, method, storage medium and computer equipment
CN112559904B (en) * 2020-12-16 2021-10-22 中国计量大学 Conversational social recommendation method based on door mechanism and multi-modal graph network
CN112883268B (en) * 2021-02-22 2022-02-01 中国计量大学 Session recommendation method considering user multiple interests and social influence
CN112905887B (en) * 2021-02-22 2021-12-14 中国计量大学 Conversation recommendation method based on multi-interest short-term priority model
CN112950325B (en) * 2021-03-16 2023-10-03 山西大学 Self-attention sequence recommendation method for social behavior fusion
CN112948683B (en) * 2021-03-16 2022-11-11 山西大学 Social recommendation method with dynamic fusion of social information
CN113407862B (en) * 2021-05-20 2022-06-17 桂林电子科技大学 Sequential social contact recommendation method, system and storage medium based on door mechanism
CN114491150B (en) * 2022-03-28 2022-07-15 苏州浪潮智能科技有限公司 Video recommendation method, system, equipment and computer readable storage medium

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101944218A (en) * 2010-01-27 2011-01-12 北京大学 Personalized recommended method based on picture under social network and system thereof
WO2018191555A1 (en) * 2017-04-14 2018-10-18 Drishti Technologies. Inc Deep learning system for real time analysis of manufacturing operations

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10803391B2 (en) * 2015-07-29 2020-10-13 Google Llc Modeling personal entities on a mobile device using embeddings

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101944218A (en) * 2010-01-27 2011-01-12 北京大学 Personalized recommended method based on picture under social network and system thereof
WO2018191555A1 (en) * 2017-04-14 2018-10-18 Drishti Technologies. Inc Deep learning system for real time analysis of manufacturing operations

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
"基于用户社交信息和搭配知识融合的服饰推荐算法研究";孙广路;《中国博士学位论文全文数据库信息科技辑》;20181115(第11期);全文 *

Also Published As

Publication number Publication date
CN109767301A (en) 2019-05-17

Similar Documents

Publication Publication Date Title
CN109767301B (en) Recommendation method and system, computer device and computer readable storage medium
CN111222054B (en) Session social contact recommendation method based on context neighbor relation modeling
Chen et al. Edge computing resources reservation in vehicular networks: A meta-learning approach
Bobadilla et al. Collaborative filtering adapted to recommender systems of e-learning
Schoder Forecasting the success of telecommunication services in the presence of network effects
Jiang et al. Modeling customer satisfaction for new product development using a PSO-based ANFIS approach
CN112733967A (en) Model training method, device, equipment and storage medium for federal learning
Yuan et al. Persistency of excitation and performance of deterministic learning
CN111125537A (en) Session recommendation method based on graph representation
CN112685657B (en) Conversation social recommendation method based on multi-mode cross fusion graph network
CN115270001B (en) Privacy protection recommendation method and system based on cloud collaborative learning
CN112528165A (en) Session social recommendation method based on dynamic routing graph network
CN114595396A (en) Sequence recommendation method and system based on federal learning
Yin et al. Near optimal provable uniform convergence in off-policy evaluation for reinforcement learning
KR20110096488A (en) Collaborative networking with optimized inter-domain information quality assessment
CN112559904B (en) Conversational social recommendation method based on door mechanism and multi-modal graph network
CN112819575B (en) Session recommendation method considering repeated purchasing behavior
Zhou et al. Towards real time team optimization
CN113449176A (en) Recommendation method and device based on knowledge graph
CN116049544A (en) Global factorization multi-feature fusion network interest prediction method
CN115600009A (en) Deep reinforcement learning-based recommendation method considering future preference of user
CN115599990A (en) Knowledge perception and deep reinforcement learning combined cross-domain recommendation method and system
CN113761388A (en) Recommendation method and device, electronic equipment and storage medium
Mönch et al. Directed Acyclic Graph-Type Distributed Ledgers via Young-Age Preferential Attachment
Wan et al. A learner’s role-based multi dimensional collaborative recommendation (LRMDCR) for group learning support

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant