CN114692005B - Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device - Google Patents

Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device Download PDF

Info

Publication number
CN114692005B
CN114692005B CN202210604217.8A CN202210604217A CN114692005B CN 114692005 B CN114692005 B CN 114692005B CN 202210604217 A CN202210604217 A CN 202210604217A CN 114692005 B CN114692005 B CN 114692005B
Authority
CN
China
Prior art keywords
sequence
user
commodity
sparse
learning
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
CN202210604217.8A
Other languages
Chinese (zh)
Other versions
CN114692005A (en
Inventor
汤胤
李泽峥
沈子璐
陈永健
路婕
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.)
Jinan University
Original Assignee
Jinan 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 Jinan University filed Critical Jinan University
Priority to CN202210604217.8A priority Critical patent/CN114692005B/en
Publication of CN114692005A publication Critical patent/CN114692005A/en
Application granted granted Critical
Publication of CN114692005B publication Critical patent/CN114692005B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0201—Market modelling; Market analysis; Collecting market data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/06—Buying, selling or leasing transactions
    • G06Q30/0601—Electronic shopping [e-shopping]
    • G06Q30/0631—Item recommendations

Abstract

The invention discloses a sparse ultrashort sequence-oriented personalized recommendation method, a system, a medium and equipment, wherein the method comprises the following steps: constructing a sparse ultrashort user behavior sequence, constructing a relational graph of users and commodities, and performing embedded representation learning on nodes in the relational graph by adopting a graph embedding method; constructing an expert database based on sparse ultra-short user behavior data, learning a purchase strategy of the expert database, and performing sparse ultra-short user behavior sequence expansion according to the purchase strategy; based on the commodity pre-embedded representation and the expanded sparse ultra-short user behavior sequence, the information enhancement of the commodity embedded representation is completed by adopting a self-attention model, the user embedded representation is obtained according to the final commodity embedded representation and the user behavior data of the sparse ultra-short user behavior sequence, and personalized recommendation is performed. The method can improve the quality of the data input into the self-attention model and realize the application of the self-attention model in a sparse ultra-short data recommendation-oriented scene.

Description

Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device
Technical Field
The invention relates to the technical field of data customization processing, in particular to a sparse ultrashort sequence-oriented personalized recommendation method, a sparse ultrashort sequence-oriented personalized recommendation system, a sparse ultrashort sequence-oriented personalized recommendation medium and sparse ultrashort sequence-oriented personalized recommendation equipment.
Background
With the development of online business and big data, the application value of the recommendation system on business intelligence is more and more obvious. The personalized recommendation method can better mine the potential interests of the user and make recommendations in a targeted manner. At present, a common practice is to recommend corresponding commodities to a user according to user characteristics, commodity characteristics and interaction information between the user and the commodities, and massive data needs to occupy a large amount of storage space. Meanwhile, in many application scenarios, in the face of an ultra-large commodity set and limited user commodity interaction times, the user commodity interaction data is in a sparse characteristic, and a new user and a low-activity user bring a cold start problem to a recommendation system, so that a commodity purchasing sequence of the user is in an ultra-short characteristic.
With the progress of artificial intelligence technology in recent years, a large-parameter complex model has the potential of better mining the potential interest of a user, a self-attention model can well capture the long-distance dependency of an object from user behavior data, but a large number of parameters are needed to learn a global structure, and the learning on a sparse and ultrashort data set is difficult.
Disclosure of Invention
In order to overcome the defects and shortcomings in the prior art, the invention provides a sparse ultrashort sequence-oriented personalized recommendation method.
The second purpose of the invention is to provide a sparse ultrashort sequence-oriented personalized recommendation system.
A third object of the present invention is to provide a computer-readable storage medium.
It is a fourth object of the invention to provide a computing device.
In order to achieve the purpose, the invention adopts the following technical scheme:
the invention provides a sparse ultrashort sequence-oriented personalized recommendation method, which comprises the following steps:
acquiring sparse ultrashort user behavior data, constructing a sparse ultrashort user behavior sequence, constructing a relation graph of a user and a commodity according to historical data of interaction between the user and the commodity, and performing embedding representation learning on nodes in the relation graph by adopting a graph embedding method;
constructing an expert database based on sparse ultrashort user behavior data, learning a purchase strategy of the expert database by adopting a simulation learning method, and completing the expansion of a sparse ultrashort user behavior sequence by following the purchase strategy;
learning the purchasing strategy of the expert database by adopting a simulated learning method, which specifically comprises the following steps:
setting a sequence length threshold, dividing a sparse ultrashort user behavior sequence into a long sequence and a short sequence according to the sequence length threshold, and storing the long sequence into an expert database;
learning the purchasing strategy of the expert database by adopting a simulated learning method based on a generated countermeasure network;
sampling a real purchase sequence state s from a sparse ultra-short user behavior sequence, and obtaining a purchase decision a by utilizing an initialized purchase strategy pi to obtain a generation experience (s, a);
sampling a long sequence from an expert database, segmenting a part of the long sequence containing the first m commodities as a state S, and taking any one of the rest commodities as an A to obtain expert experience (S, A);
the production experience (S, a) and the expert experience (S, a) are simultaneously input into the discriminator D, and the difference is calculated using the cross entropy:
Figure 197339DEST_PATH_IMAGE001
wherein the content of the first and second substances,
Figure 924993DEST_PATH_IMAGE002
the result of the calculation of the difference is represented,
Figure 240567DEST_PATH_IMAGE003
for true purchase sequence statussDown to policy selection action
Figure 25334DEST_PATH_IMAGE004
The probability of (a) of (b) being,
Figure 878889DEST_PATH_IMAGE005
for the score expectation of the purchase strategy pi,
Figure 895387DEST_PATH_IMAGE006
to usePolicy
Figure 517998DEST_PATH_IMAGE007
The score expectation of (a) is that,
Figure 211017DEST_PATH_IMAGE008
an intermediate parameter is expressed for measuring the uncertainty of the purchasing strategy pi,
Figure 99338DEST_PATH_IMAGE009
representing a preset weight parameter;
iteratively updating the weight parameters of the purchasing strategy pi, learning to obtain the purchasing strategy pi and a reward function, inputting the state s of the real purchasing short sequence into the deep neural network of the purchasing strategy pi and obtaining a purchasing decisionaWill purchase the decisionaAdding real purchase short sequences to complete length expansion until the length is expanded to a preset length;
updating a purchasing strategy pi by taking the imitation learning as a framework;
based on the commodity pre-embedded representation and the expanded sparse ultra-short user behavior sequence, the information enhancement of the commodity embedded representation is completed by adopting a self-attention model, the user embedded representation is obtained according to the final commodity embedded representation and the user behavior data of the sparse ultra-short user behavior sequence, and personalized recommendation is performed according to the embedded representation of the commodity and the user.
As a preferred technical scheme, a graph embedding method is adopted to carry out embedding representation learning on nodes in a relational graph, and the method specifically comprises the following steps:
constructing a user and commodity relation bipartite graph according to a commodity purchasing history sequence of a user;
generating a commodity purchasing sequence of the user and a commodity purchasing sequence of the user, and constructing an inducing lists x And inducement listss y ；
Dividing a user-commodity relationship bipartite graph into a plurality of induction listss x Ands y the embedded learning is designed as a maximum likelihood problem;
initializing user purchase goods prediction network N x Commodity purchase prediction network N y Attribute-embedded network N a Initializing embedded representation of users or commodities required by an embedding layer, and initializing a weight matrix required by a softmax layer;
from the list of inducerss x Extract an edge from the induction lists x Sampling other edges as the neighbors;
from the list of inducerss y Extract an edge from the induction lists y Sampling other edges as the neighbors;
setting update target, updating prediction network N x Predictive network N y Attribute-embedded network N a And a commodity or user embedded representation vocabulary;
training until convergence, and updating the weight matrix parameters to obtain the embedded representation of the commodity and the user.
As a preferred technical solution, the construction of the bipartite graph of the relationship between the user and the product according to the history sequence of the product purchased by the user is specifically represented as: g (X, Y, A, t), wherein X, Y respectively represents the user node and the commodity node set, A represents the relevant attribute of the user interacting with the commodity, and t represents the time stamp of the interaction.
As a preferred technical scheme, the induction list is constructeds x And inducement lists y The method comprises the following specific steps:
generating a sequence x of purchases of goods by a user node x * =[(Y * ,A,t)]And the sequence y of the purchase of the goods y by the user * =[(X * ,A,t)]Wherein X is * To purchase a set of users of Y goods, Y * Obtaining a guidance list for the commodity set purchased by the x user nodes according to the purchase sequence
Figure 171724DEST_PATH_IMAGE010
And inducement lists
Figure 727470DEST_PATH_IMAGE011
Wherein, in the step (A),
Figure 122548DEST_PATH_IMAGE012
is x * And Y * The established connecting edge is connected with the network,
Figure 763745DEST_PATH_IMAGE013
is y and X * And C, establishing a connecting edge, wherein A represents the related attribute of the interaction between the user and the commodity, and t represents the time stamp of the occurrence of the interaction.
Preferably, the bipartite graph of relationships between users and commodities is divided into a plurality of inducement listss x Ands y the embedded learning is designed as a maximum likelihood problem, which is specifically expressed as:
Figure 653204DEST_PATH_IMAGE014
wherein the content of the first and second substances,
Figure 984828DEST_PATH_IMAGE015
is a balance parameter preset according to experience and is used for controlling the balance of the oil in the oil tanks x Ands y is to be weighed against the importance of,
Figure 19649DEST_PATH_IMAGE016
are all the parameters that are involved in the model,
Figure 66584DEST_PATH_IMAGE017
representing an objective function;
computing
Figure 477842DEST_PATH_IMAGE018
Specifically, it is represented as:
Figure 8181DEST_PATH_IMAGE019
computing
Figure 682745DEST_PATH_IMAGE020
Specifically, it is represented as:
Figure 219905DEST_PATH_IMAGE021
wherein the content of the first and second substances,
Figure 375468DEST_PATH_IMAGE022
and
Figure 494733DEST_PATH_IMAGE023
respectively represent x andy i an embedded representation of a node is shown,
Figure 402515DEST_PATH_IMAGE022
an embedding matrix from the set of user nodes X,
Figure 240021DEST_PATH_IMAGE024
an embedding matrix from the commodity node set Y, the computation of the embedding matrix taking into accounty i Properties of nodesd i ，d i Is an attribute embedded network N a The output result of (a) is obtained,bis the term of the offset, and,
Figure 9394DEST_PATH_IMAGE025
as softmax weighting matrices
Figure 763592DEST_PATH_IMAGE026
The k-th row of the matrix is transposed,
Figure 327429DEST_PATH_IMAGE027
as softmax weighting matrices
Figure 104761DEST_PATH_IMAGE028
First of the matrixlThe lines are transposed so that the lines are transposed,
Figure 393004DEST_PATH_IMAGE029
is taking into account the attributesd j Softmax weight matrix of
Figure 486862DEST_PATH_IMAGE030
The line k of (a) is transposed,
Figure 939709DEST_PATH_IMAGE031
is taking into account the attributesd j Softmax weight matrix of
Figure 486228DEST_PATH_IMAGE032
To (1) alThe lines are transposed so that the lines are transposed,
Figure 581092DEST_PATH_IMAGE033
is taking into account the attributed j Softmax weight matrix of
Figure 529456DEST_PATH_IMAGE034
The line k of (a) is transposed,
Figure 684363DEST_PATH_IMAGE035
is taking into account the attributesd j Softmax weight matrix of
Figure 452599DEST_PATH_IMAGE034
To (1) alThe lines are transposed so that the lines are transposed,
Figure 354083DEST_PATH_IMAGE036
is a preset hyper-parameter, representing the number of neighbors,a i represents x andy i the associated attributes of the interaction of (a),a j represents x andy j x represents a user node,y i representing a commodityiThe node of (a) is selected,y j representing a commodityjA node of (2);
v y andv xi respectively representyAndx i an embedded representation of a node is shown,v y an embedding matrix from the commodity node set Y,
Figure 156954DEST_PATH_IMAGE037
an embedding matrix from the set of user nodes X, the computation of which takes into accountx i Properties of nodesd i ，
Figure 748341DEST_PATH_IMAGE038
As softmax weight matrix
Figure 535032DEST_PATH_IMAGE039
The transpose of the k-th line of (1),
Figure 237278DEST_PATH_IMAGE040
as softmax weight matrix
Figure 160234DEST_PATH_IMAGE041
To (1) alAnd (5) line transposition.
As a preferred technical solution, the update weight matrix parameter is expressed as:
Figure 656943DEST_PATH_IMAGE042
wherein the content of the first and second substances,
Figure 930930DEST_PATH_IMAGE043
represents iteration totThe parameters of the model at the time of generation,
Figure 187599DEST_PATH_IMAGE044
is the direction of decrease of the parameter,
Figure 234837DEST_PATH_IMAGE045
the step size is represented as a function of time,
Figure 184339DEST_PATH_IMAGE015
are balance parameters that are preset empirically.
As a preferred technical solution, the information enhancement of the commodity embedded representation is completed by using a self-attention model, the user embedded representation is obtained according to the final commodity embedded representation and the user behavior data of the sparse ultrashort user behavior sequence, and personalized recommendation is performed according to the embedded representations of the commodity and the user, and the specific steps include:
inputting a real commodity purchasing length sequence of a user and a commodity purchasing length sequence of a virtual user after the short sequence is expanded in a self-attention model;
performing mask preprocessing on the training samples;
training a self-attention model to be convergent, inputting a historical sequence of commodities purchased by each user for the model to obtain an embedded representation of the commodities appearing in the sequence, and aggregating different embedded representations of the same commodity by adopting a mean aggregation or weighted aggregation mode to obtain a final embedded representation of the commodity;
obtaining user embedded representation by adopting a mean value aggregation commodity final embedded representation mode according to a historical commodity purchasing sequence of a real user;
and calculating and sequencing the cosine similarity of the users and the commodities, and recommending the commodity with the highest similarity to each user.
In order to achieve the second object, the invention adopts the following technical scheme:
a sparse ultrashort sequence-oriented personalized recommendation system comprises: the system comprises a user behavior sequence building module, a relational graph building module, an embedded representation learning module, an expert database building module, an imitation learning module, a sequence expansion module, an embedded representation output module and a personalized recommendation module;
the user behavior sequence construction module is used for acquiring sparse ultrashort user behavior data and constructing a sparse ultrashort user behavior sequence;
the relation graph building module is used for building a relation graph of the user and the commodity according to historical data of interaction between the user and the commodity;
the embedded representation learning module is used for embedding representation learning of the nodes in the relational graph by adopting a graph embedding method;
the expert database construction module is used for constructing an expert database based on sparse ultra-short user behavior data;
the imitation learning module is used for learning the purchasing strategy of the expert database by adopting an imitation learning method;
learning the purchasing strategy of the expert database by adopting a simulated learning method, which specifically comprises the following steps:
setting a sequence length threshold, dividing a sparse ultrashort user behavior sequence into a long sequence and a short sequence according to the sequence length threshold, and storing the long sequence into an expert database;
learning the purchasing strategy of the expert database by adopting a simulated learning method based on a generated countermeasure network;
sampling a real purchase sequence state s from a sparse ultra-short user behavior sequence, and obtaining a purchase decision a by utilizing an initialized purchase strategy pi to obtain a generation experience (s, a);
sampling a long sequence from an expert database, segmenting a part of the long sequence containing the first m commodities as a state S, and taking any one of the rest commodities as an A to obtain expert experience (S, A);
the production experience (S, a) and the expert experience (S, a) are simultaneously input into the discriminator D, and the difference is calculated using the cross entropy:
Figure 929310DEST_PATH_IMAGE001
wherein, the first and the second end of the pipe are connected with each other,
Figure 724091DEST_PATH_IMAGE002
the result of the calculation of the difference is represented,
Figure 621639DEST_PATH_IMAGE003
for true purchase sequence statussSelecting actions according to policy
Figure 991310DEST_PATH_IMAGE004
The probability of (a) of (b) being,
Figure 974309DEST_PATH_IMAGE005
for the score expectation of the purchase strategy pi,
Figure 572781DEST_PATH_IMAGE006
for using the strategy
Figure 839683DEST_PATH_IMAGE007
The score expectation of (a) is,
Figure 865408DEST_PATH_IMAGE008
an intermediate parameter is expressed for measuring the uncertainty of the purchasing strategy pi,
Figure 587901DEST_PATH_IMAGE009
representing a preset weight parameter;
iteratively updating the weight parameter of the purchasing strategy pi, learning to obtain the purchasing strategy pi and a reward function, inputting the state s of the real purchasing short sequence into the deep neural network of the purchasing strategy pi and obtaining a purchasing decisionaWill purchase the decisionaAdding real purchase short sequences to complete length expansion until the length is expanded to a preset length;
updating a purchasing strategy pi by taking the simulation learning as a framework;
the sequence expansion module is used for completing the expansion of the sparse ultrashort user behavior sequence by adopting a purchase strategy;
the embedded representation output module is used for finishing information enhancement of commodity embedded representation by adopting a self-attention model based on commodity pre-embedded representation and the expanded sparse ultra-short user behavior sequence, and obtaining user embedded representation according to the final commodity embedded representation and the user behavior data of the sparse ultra-short user behavior sequence;
the personalized recommendation module is used for performing personalized recommendation according to the embedded representations of the commodities and the users.
In order to achieve the third object, the invention adopts the following technical scheme:
a computer-readable storage medium, storing a program, which when executed by a processor, implements the sparse ultrashort sequence-oriented personalized recommendation method as described above.
In order to achieve the fourth object, the invention adopts the following technical scheme:
a computer device comprises a processor and a memory for storing a program executable by the processor, wherein the processor executes the program stored in the memory to realize the sparse ultrashort sequence-oriented personalized recommendation method.
Compared with the prior art, the invention has the following advantages and beneficial effects:
(1) according to the method, the generated graph is embedded into the pre-representation, the data sequence is expanded, deep learning of the self-attention module learning embedding representation effect is enhanced, and the self-attention model is applied to a scene for recommending sparse ultra-short sequence data, so that the mining capability of potential interest of a user is enhanced, and the personalized recommendation accuracy is improved.
(2) The method distinguishes the user behavior sequence into the long sequence and the short sequence, learns the purchase strategy which is easier to learn in the long sequence by using the imitation learning so as to expand the short sequence to the specified length, solves the problem of sparse and ultrashort quality of a training sample, improves the sample quality of an input self-attention model, depends on massive sparse ultrashort data, and occupies small storage space.
(3) The invention describes the commodity relation of the user by using a graph structure, learns the commodity embedded representation by using a graph embedding method to obtain high-quality commodity pre-embedded representation and improve the pre-representation quality of the input self-attention model sample.
Drawings
FIG. 1 is a schematic flow chart of a sparse ultrashort sequence-oriented personalized recommendation method according to the present invention;
FIG. 2 is a schematic diagram of an extended sequence according to the present invention;
FIG. 3 is a schematic diagram of a node-embedded representation of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Example 1
As shown in fig. 1, the present embodiment provides a sparse ultrashort sequence-oriented personalized recommendation method, including the following steps:
s1, acquiring sparse and ultrashort user behavior data from a storage space, constructing a user-commodity relation graph based on the sparse and ultrashort user behavior data, and completing pre-embedding expression of commodities by adopting a graph embedding method;
in step S1, the pre-embedding of the merchandise is completed by the graph embedding method, which includes the following steps;
s11, constructing a graph structure G according to the historical sequence of the commodities purchased by the user;
s12, based on the graph structure G, embedding representation learning is carried out on the nodes in the graph by adopting a graph embedding method;
preferably, in the recommendation problem, a user commodity relation is expressed by adopting a heterogeneous graph structure;
preferably, a heterogeneous graph (such as a bipartite graph) embedding method is used for representing and learning nodes, the relationship between the heterogeneous graph and the commodity can be described by the heterogeneous graph (such as the bipartite graph) according to historical data of interaction between users and commodities, a user commodity relationship bipartite graph G = (X, Y, E) is given, X, Y respectively represents node sets of users and commodities, E represents a connecting edge of nodes in the node sets of X and Y, after the relationship bipartite graph is constructed, node embedding is learned by using a deep neural network, and the goal of node embedding is to obtain a function
Figure 990063DEST_PATH_IMAGE046
Sum function
Figure 862204DEST_PATH_IMAGE047
，
Figure 308098DEST_PATH_IMAGE048
And
Figure 265690DEST_PATH_IMAGE049
are bipartite graph nodes respectively
Figure 924073DEST_PATH_IMAGE050
K and
Figure 650721DEST_PATH_IMAGE051
performing dimensional characterization;
in this embodiment, the bipartite graph embedding adopts an IGE (attribute interaction graph) embedding method, which specifically includes the following steps:
(1) constructing a bipartite graph G (X, Y, A, t) according to a commodity purchasing history sequence of a user, wherein X and Y are a user node and commodity node set, A is relevant attributes of interaction (purchase) between the user and commodities, such as purchase price, commodity category and the like, and t is a time stamp of interaction occurrence;
(2) generating a sequence x of purchases of goods by a user node x * =[(Y * ,A,t)]And the sequence y of the purchase of the goods y by the user * =[(X * ,A,t)]Wherein X is * To purchase a set of users of Y goods, Y * Obtaining a guidance list according to the purchase sequence for the commodity set purchased by the x user nodes
Figure 267516DEST_PATH_IMAGE052
And inducement lists
Figure 712403DEST_PATH_IMAGE011
Wherein, in the step (A),
Figure 190789DEST_PATH_IMAGE012
is x * And Y * Established connecting edge
Figure 752702DEST_PATH_IMAGE053
，
Figure 291131DEST_PATH_IMAGE054
Is y and X * Built connecting edge, lists x /s y Contains information of all edges related to the node x/y;
(3) embedded learning can be designed as a maximum likelihood problem, and graph G can be divided into multiple inducement listss x Ands y then the maximum likelihood function is shown in equation 1:
Figure 472582DEST_PATH_IMAGE014
(formula 1)
Wherein the content of the first and second substances,
Figure 489080DEST_PATH_IMAGE015
is a balance parameter preset according to experience and is used for controlling the balance of the oil in the oil tanks x Ands y is to be weighed against the importance of,
Figure 455899DEST_PATH_IMAGE016
is all parameters related to the model, the objective function (formula 1) has two similar parts, steps (5) - (8) describe the calculation of the first part of the objective function, and steps (9) - (12) describe the calculation of the second part of the objective function;
(4) initializing user purchased goods prediction network N x Commodity purchasing prediction network N y Attribute-embedded network N a The embedded representation of the user or commodity required to initialize the embedding layer, i.e. initializing the weight matrix
Figure 414496DEST_PATH_IMAGE055
Figure 568397DEST_PATH_IMAGE056
WhereinFIs a pre-selected number of factors that,V x is the size of the set of user nodes,V y is the size of the node set of the commodity,Dis the dimension to which the attribute network is reduced,Kis the embedding dimension of the user and,
Figure 169012DEST_PATH_IMAGE051
is the embedding dimension of the commodity, and a weight matrix required by initializing the softmax layer
Figure 724758DEST_PATH_IMAGE057
，
Figure 870568DEST_PATH_IMAGE058
(5) Froms x Extract a side
Figure 763963DEST_PATH_IMAGE059
And froms x Middle sampling
Figure 387842DEST_PATH_IMAGE060
Other edges of the strip
Figure 63674DEST_PATH_IMAGE061
As its neighbor, sample to anythinge j Probability of and
Figure 629653DEST_PATH_IMAGE062
is proportional to the size of the (c), wherein,
Figure 758146DEST_PATH_IMAGE063
is an adjustable hyper-parameter to take into account samplinge j The effect of the time of the (c) phase,
Figure 434984DEST_PATH_IMAGE036
representing preset hyper-parameters, such as fixing that each user node has several commodity nodes as neighbors,xto indicate the user or users of the device,y i indicating merchandiseiThe node(s) of (a) is (are),y j indicating merchandisejThe node(s) of (a) is (are),t i represents x andy i at the interaction oft i The time of day (e.g. a date),t j represents x andy j at the interaction oft i At the moment of time, the time of day,a i represents x andy i related attributes of the interaction (e.g., x purchase)y i The closing price of (c),a j represents x andy j the relevant attributes of the interaction of (1);
(6) according to equation 2
Figure 699743DEST_PATH_IMAGE064
The following are:
Figure 187357DEST_PATH_IMAGE065
(formula 2)
Wherein the content of the first and second substances,y j =kcan be understood asy k ，
Figure 786834DEST_PATH_IMAGE066
And
Figure 18095DEST_PATH_IMAGE023
respectively represent x andy i an embedded representation of a node is shown,
Figure 649278DEST_PATH_IMAGE066
embedding matrix from user node set X
Figure 42213DEST_PATH_IMAGE067
，
Figure 145299DEST_PATH_IMAGE024
Embedding matrix from commodity node set Y
Figure 163939DEST_PATH_IMAGE068
Figure 668870DEST_PATH_IMAGE069
，d i Is an attribute embedded network N a The output result of (a) is obtained,
Figure 232706DEST_PATH_IMAGE070
is the term of the offset, and,
Figure 72355DEST_PATH_IMAGE025
can be understood as
Figure 645419DEST_PATH_IMAGE026
The transpose of the k-th row of the matrix,
Figure 988544DEST_PATH_IMAGE027
as softmax weighting matrices
Figure 988861DEST_PATH_IMAGE028
First of the matrixlThe lines are transposed to form a line,
Figure 800960DEST_PATH_IMAGE029
is taking into account the attributesd j Softmax weight matrix of
Figure 429912DEST_PATH_IMAGE030
The line k of (a) is transposed,
Figure 643855DEST_PATH_IMAGE031
is taking into account the attributesd j Softmax weight matrix of
Figure 815074DEST_PATH_IMAGE032
To (1) alLine transpose, the calculation of which is as shown in step (4);
(7) according to equation 3
Figure 363735DEST_PATH_IMAGE018
Figure 278602DEST_PATH_IMAGE071
(formula 3)
(8) To be provided with
Figure 596320DEST_PATH_IMAGE072
As large as possible for the target calculation
Figure 938439DEST_PATH_IMAGE073
Updating the prediction network N with Adam according to equation 4 x Attribute embedded network N a And commodity/user embedded representation vocabulary (i.e., parameter matrix initialized by the embedding layer), wherein,
Figure 459550DEST_PATH_IMAGE043
represents iteration totThe parameters of the model at the time of generation,
Figure 427375DEST_PATH_IMAGE073
is the direction of decrease of the parameter,
Figure 615911DEST_PATH_IMAGE045
is the step size;
Figure 128932DEST_PATH_IMAGE042
(formula 4)
(9) Froms y Extract a side
Figure 649257DEST_PATH_IMAGE059
And froms y Middle sampling
Figure 905926DEST_PATH_IMAGE060
Other edges of the strip
Figure 198235DEST_PATH_IMAGE061
As its neighbor, sample to anythinge j Probability of and
Figure 882158DEST_PATH_IMAGE062
is proportional to the size of the (c), wherein,
Figure 377861DEST_PATH_IMAGE063
is an adjustable hyper-parameter to take into account samplinge j The time effect of (a);
(10) according to equation 5
Figure 687489DEST_PATH_IMAGE074
The following are:
Figure 585037DEST_PATH_IMAGE075
(formula 5)
Wherein the content of the first and second substances,x j =kcan be understood asx k ，v y Andv xi is composed ofyAndx i an embedded representation of a node is shown,v y embedded matrix from Y
Figure 439861DEST_PATH_IMAGE076
，
Figure 203286DEST_PATH_IMAGE037
Embedded matrix from X
Figure 801758DEST_PATH_IMAGE077
Figure 553813DEST_PATH_IMAGE078
，
Figure 831735DEST_PATH_IMAGE079
Is an attribute embedded network N a The output result of (a) is obtained,
Figure 302031DEST_PATH_IMAGE080
is the term of the offset, and,
Figure 219040DEST_PATH_IMAGE081
can be understood as
Figure 560023DEST_PATH_IMAGE082
The transpose of the k-th row of the matrix,
Figure 287808DEST_PATH_IMAGE033
is taking into account the attributesd j Softmax weight matrix of
Figure 229088DEST_PATH_IMAGE034
The line k of (a) is transposed,
Figure 169362DEST_PATH_IMAGE035
is taking into account the attributed j Softmax weight matrix of
Figure 161589DEST_PATH_IMAGE034
To (1) alThe lines are transposed so that the lines are transposed,
Figure 778384DEST_PATH_IMAGE036
is a preset hyper-parameter, representing the number of neighbors,
Figure 957692DEST_PATH_IMAGE038
as softmax weight matrix
Figure 783933DEST_PATH_IMAGE039
The transpose of the k-th line of (1),
Figure 630666DEST_PATH_IMAGE040
as softmax weight matrix
Figure 169095DEST_PATH_IMAGE041
To (1) alAnd (5) line transposition. The calculation is as shown in step (4);
(11) according to equation 6
Figure 350546DEST_PATH_IMAGE083
Figure 632623DEST_PATH_IMAGE084
(formula 6)
(12) To be provided with
Figure 583130DEST_PATH_IMAGE085
As large as possible for the target calculation
Figure 292460DEST_PATH_IMAGE044
Updating the prediction network N with Adam according to equation 7 y Attribute-embedded network N a And commodity/user embedded expression vocabulary (i.e. parameter matrix initialized by embedding layer), wherein
Figure 711940DEST_PATH_IMAGE043
Represents iteration totThe parameters of the model at the time of generation,
Figure 46976DEST_PATH_IMAGE044
is the direction of decrease of the parameter,
Figure 602722DEST_PATH_IMAGE045
represents a step size;
Figure 748532DEST_PATH_IMAGE042
(formula 4)
(13) Training the steps (5) - (8) and the steps (9) - (12) alternately until convergence, and updating the weight matrix parameters to obtain the embedded representation of the commodity and the user;
s2, constructing an expert database based on sparse ultra-short user behavior data, learning the purchase strategy of the expert database by adopting a simulation learning method, and completing the expansion of a user sparse ultra-short purchase sequence by following the purchase strategy;
as shown in fig. 2, in step S2, the steps of the mimic learning method adopted by the expansion sequence are as follows;
s21, dividing the sparse and ultrashort user behavior sequence into a long sequence and a short sequence, and storing the long sequence into an expert database;
specifically, a sequence length threshold L (set depending on the particular data set) may be set, with sequence length ≧ L being long sequence, and sequence length < L being short sequence;
s22, learning the purchase strategy from the expert database and expanding the sparse ultrashort sequence accordingly;
preferably, the purchasing strategy is learned from an expert database, the sparse ultrashort sequence is expanded accordingly, and a simulation learning method based on a generation countermeasure network is adopted, wherein the steps are as follows;
s221, sampling a real purchasing sequence state S from a user behavior sequence, obtaining a purchasing decision a by utilizing an initialized purchasing strategy pi, and accordingly obtaining a generating experience (S, a);
the real purchase sequence state s is a weighted average value of historical purchased commodity embedded expressions before the purchase decision is made, the weight of the real purchase sequence state s is a learnable parameter, a candidate set of the purchase decision is all purchasable commodities, and a purchase strategy pi and a discriminator D are initialized deep neural networks;
s222, sampling a long sequence from an expert database, segmenting a part of the long sequence containing the first m commodities as a state S, and taking any one of the rest commodities as an A to obtain expert experience (S, A);
s223, simultaneously inputting the generated experience (S, a) and the expert experience (S, a) into the discriminator D to obtain scores D (S, a) and D (S, a) of the two, when the decision-making discrimination score (reward) is the difference between the two, see formula 4, which shows that the difference between the two can be expressed as the sum of the difference between D (S, a) and the real target score (e.g. 1) and the difference between D (S, a) and the false target score (e.g. 0) when implemented, and the difference can be calculated using cross entropy:
Figure 641927DEST_PATH_IMAGE086
(formula 7)
Wherein the content of the first and second substances,
Figure 265806DEST_PATH_IMAGE087
to measure the uncertainty of the purchasing strategy pi,
Figure 190906DEST_PATH_IMAGE003
for true purchase sequence statussSelecting actions according to policy
Figure 242038DEST_PATH_IMAGE004
The probability of (a) of (b) being,
Figure 370531DEST_PATH_IMAGE005
for the score expectation of the purchase strategy pi, the formula for calculating the score is a formula in middle brackets, the weight parameter of the discriminator D is updated by Adam so that the formula 7 is as large as possible, and the weight parameter theta of the purchase strategy pi is updated by TRPO so that the formula 7 is as large as possible
Figure 47369DEST_PATH_IMAGE088
As small as possible in the form of a capsule,
Figure 577707DEST_PATH_IMAGE006
to use the strategy
Figure 65321DEST_PATH_IMAGE007
The score expectation of (a) is,
Figure 930377DEST_PATH_IMAGE009
representing a preset weight parameter;
s224, continuously repeating the process, and finally learning out a pi meeting the expert experience E On the basis of which the true purchases can be ordered in short sequences (of length m) and a reward function 1 ) State of (1)s*Inputting a deep neural network of a purchasing strategy pi and obtaining a purchasing decisionaAdding the purchasing decision into the real short sequence to complete the expansion of one length, and adding the expanded sequence (with the length of m) 1 State of + 1)s*Repeating the above operation, thus circulating N-m 1 Expanding to a specified length N;
in the embodiment, simulation learning is taken as a framework, a countermeasure generation and deep reinforcement learning technology is adopted for realizing, and a purchasing strategy pi is updated through algorithms of TRPO, PPO or model-based and the like;
the purchase strategy pi can be updated by an algorithm of TRPO, PPO or model-based, etc.;
s3, based on the commodity pre-embedded representation and the expanded user behavior data, completing information enhancement of the commodity embedded representation by adopting a self-attention model, obtaining a user embedded representation according to the final commodity embedded representation and the real user behavior data, and performing personalized recommendation according to the embedded representation of the commodity and the user;
as shown in fig. 3, the information enhancement of the embedded representation of the commodity is completed by using the self-attention model, and the steps are as follows:
s31, providing high-quality sample input for the self-attention model, wherein the sample is pre-expressed as a result of learning by using a graph embedding method, and the problem of sparseness and ultrashort is relieved after a short sequence in the sample is subjected to sequence expansion;
s32, performing commodity representation learning on the input samples by using a self-attention model;
s33, obtaining user representation based on the user behavior data and the final commodity representation, and recommending according to the matching degree of the user and the commodity;
in this embodiment, the self-attention model is a BERT model, and the specific steps are as follows:
(1) inputting a real commodity long sequence purchased by a user and a commodity long sequence purchased by a virtual user after the short sequence is expanded for BERT, wherein the pre-embedding of each commodity in the sequence is represented as a learning result of a bipartite graph embedding model IGE;
(2) i% masking of the training samples, the masked commodity will have j% replaced, where j 1 % is replaced with random Commodity, j 2 % is replaced by a mask vector, k% of the commodities which are masked are kept unchanged, and the parameter setting is determined according to a specific data set;
(3) training a BERT model to be convergent, inputting a historical sequence of purchased commodities of each user into the model, obtaining an embedded representation of commodities appearing in the sequence, aggregating different embedded representations of the same commodity in a mean value aggregation or weighted aggregation (the weight can be positively correlated with the sequence length because a longer purchased commodity sequence has higher probability to reflect the purchasing interest of the user, and the randomness of a short sequence to reflect the user interest is high) mode and the like to obtain a final embedded representation of the commodity, and obtaining the user embedded representation in a mean value aggregated commodity final embedded representation mode according to the historical purchased commodity sequence of the real user;
in this embodiment, the bipartite graph embedding is mainly to obtain a good initial product embedding representation, and the user embedding representation here can be obtained according to the final product embedding representation, so as to improve the accuracy of similarity calculation between the final product and the user;
(4) calculating cosine similarity between the user and the commodities, sequencing from big to small, and recommending topK commodities with the highest similarity for each user;
in this embodiment, the representation of the user is the average of the goods that the user has historically purchased;
in this embodiment, the matching degree calculation function between the user and the product is cosine similarity, the embedding dimensions of the user and the product are consistent, and the embedding representation of the user is derived from the related aggregation calculation of the product embedding representation and includes product information, so the cosine similarity can be used to calculate the similarity of the user and the product.
According to the commodity pre-embedded representation obtained by the graph embedding method, the problem of data sparseness is solved by adopting a technical scheme of simulation learning, the technical scheme of a self-attention model is adopted after an expanded data set and the pre-embedded representation are obtained, the technical problem of global insufficient connection among commodities is solved, and finally the high-quality commodity pre-embedded representation is obtained.
Example 2
A sparse ultrashort sequence-oriented personalized recommendation system comprises: the system comprises a user behavior sequence building module, a relational graph building module, an embedded representation learning module, an expert database building module, an imitation learning module, a sequence expansion module, an embedded representation output module and a personalized recommendation module;
in this embodiment, the user behavior sequence construction module is configured to obtain sparse ultrashort user behavior data and construct a sparse ultrashort user behavior sequence;
in this embodiment, the relationship graph building module is configured to build a relationship graph between a user and a commodity according to historical data of interaction between the user and the commodity;
in this embodiment, the embedded representation learning module is configured to perform embedded representation learning on nodes in the relational graph by using a graph embedding method;
the method for embedding, representing and learning the nodes in the relational graph by adopting a graph embedding method specifically comprises the following steps:
constructing a user and commodity relation bipartite graph according to a commodity purchasing history sequence of a user;
generating a commodity purchasing sequence of the user and a commodity purchasing sequence of the user, and constructing an inducing lists x And inducement lists y ；
Dividing a user-commodity relationship bipartite graph into a plurality of induction listss x Ands y design the embedded learning as a maximumA likelihood problem;
initializing user purchase goods prediction network N x Commodity purchase prediction network N y Attribute-embedded network N a Initializing embedded representation of users or commodities required by an embedding layer, and initializing a weight matrix required by a softmax layer;
from the list of inducerss x Extract an edge from the induction lists x Sampling other edges as the neighbors;
from the list of inducerss y Extract an edge from the induction lists y Sampling other edges as the neighbors;
setting update target, updating prediction network N x Predictive network N y Attribute-embedded network N a And a commodity or user embedded representation vocabulary;
training until convergence, and updating the weight matrix parameters to obtain the embedded representation of the commodity and the user.
In the embodiment, the expert database construction module is used for constructing an expert database based on sparse ultra-short user behavior data;
in this embodiment, the imitation learning module is configured to learn a purchase strategy of the expert database by using an imitation learning method;
learning the purchasing strategy of the expert database by adopting a simulated learning method, which specifically comprises the following steps:
setting a sequence length threshold, dividing a sparse ultrashort user behavior sequence into a long sequence and a short sequence according to the sequence length threshold, and storing the long sequence into an expert database;
learning the purchasing strategy of the expert database by adopting a simulated learning method based on a generated countermeasure network;
sampling a real purchase sequence state s from a sparse ultra-short user behavior sequence, and obtaining a purchase decision a by utilizing an initialized purchase strategy pi to obtain a generation experience (s, a);
sampling a long sequence from an expert database, segmenting a part of the long sequence containing the first m commodities as a state S, and taking any one of the rest commodities as an A to obtain expert experience (S, A);
the production experience (S, a) and the expert experience (S, a) are simultaneously input into the discriminator D, and the difference is calculated using the cross entropy:
Figure 630480DEST_PATH_IMAGE001
wherein, the first and the second end of the pipe are connected with each other,
Figure 261663DEST_PATH_IMAGE002
the result of the calculation of the difference is represented,
Figure 920177DEST_PATH_IMAGE003
for true purchase sequence statussSelecting actions according to policy
Figure 23263DEST_PATH_IMAGE004
The probability of (a) of (b) being,
Figure 776324DEST_PATH_IMAGE005
for the score expectation of the purchase strategy pi,
Figure 15675DEST_PATH_IMAGE006
to use the strategy
Figure 828779DEST_PATH_IMAGE007
The score expectation of (a) is,
Figure 419161DEST_PATH_IMAGE009
representing a preset weight parameter;
iteratively updating the weight parameters of the purchasing strategy pi, learning to obtain the purchasing strategy pi and a reward function, inputting the state s of the real purchasing short sequence into the deep neural network of the purchasing strategy pi and obtaining a purchasing decisionaWill purchase the decisionaAdding real purchase short sequences to complete length expansion until the length is expanded to a preset length;
updating a purchasing strategy pi by taking the simulation learning as a framework;
in this embodiment, the sequence expansion module is configured to complete expansion of a sparse and ultrashort user behavior sequence by using a purchase strategy;
in this embodiment, the embedded representation output module is configured to complete information enhancement of the commodity embedded representation by using a self-attention model based on the commodity pre-embedded representation and the extended sparse ultra-short user behavior sequence, and obtain the user embedded representation according to the final commodity embedded representation and the user behavior data of the sparse ultra-short user behavior sequence;
in this embodiment, the personalized recommendation module is configured to perform personalized recommendation according to the embedded representations of the goods and the user.
Example 3
The present embodiment provides a storage medium, which may be a storage medium such as a ROM, a RAM, a magnetic disk, an optical disk, or the like, where one or more programs are stored, and when the programs are executed by a processor, the method for personalized recommendation based on sparse ultrashort sequence oriented in embodiment 1 is implemented.
Example 4
The embodiment provides a computing device, which may be a desktop computer, a notebook computer, a smart phone, a PDA handheld terminal, a tablet computer, or other terminal devices with a display function, and the computing device includes a processor and a memory, where the memory stores one or more programs, and when the processor executes the programs stored in the memory, the sparse ultrashort sequence-oriented personalized recommendation method in embodiment 1 is implemented.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims (10)

1. A sparse ultrashort sequence-oriented personalized recommendation method is characterized by comprising the following steps:
acquiring sparse ultrashort user behavior data, constructing a sparse ultrashort user behavior sequence, constructing a relation graph of a user and a commodity according to historical data of interaction between the user and the commodity, and performing embedding representation learning on nodes in the relation graph by adopting a graph embedding method;
constructing an expert database based on sparse ultrashort user behavior data, learning a purchase strategy of the expert database by adopting a simulation learning method, and completing the expansion of a sparse ultrashort user behavior sequence by following the purchase strategy;
learning the purchasing strategy of the expert database by adopting a simulated learning method, which specifically comprises the following steps:
setting a sequence length threshold, dividing a sparse ultrashort user behavior sequence into a long sequence and a short sequence according to the sequence length threshold, and storing the long sequence into an expert database;
learning the purchasing strategy of the expert database by adopting a simulated learning method based on a generated countermeasure network;
sampling a real purchase sequence state s from a sparse ultra-short user behavior sequence, and obtaining a purchase decision a by utilizing an initialized purchase strategy pi to obtain a generation experience (s, a);
sampling a long sequence from an expert database, segmenting a part of the long sequence containing the first m commodities as a state S, and taking any one of the rest commodities as an A to obtain expert experience (S, A);
the production experience (S, a) and the expert experience (S, a) are simultaneously input into the discriminator D, and the difference is calculated using the cross entropy:
Figure 903454DEST_PATH_IMAGE002
wherein the content of the first and second substances,
Figure 555015DEST_PATH_IMAGE004
the result of the calculation of the difference is represented,
Figure 256124DEST_PATH_IMAGE006
for true purchase sequence statussDown to policy selection action
Figure 385754DEST_PATH_IMAGE008
The probability of (a) of (b) being,
Figure 124427DEST_PATH_IMAGE010
for the score expectation of the purchase strategy pi,
Figure 681311DEST_PATH_IMAGE012
to use the strategy
Figure 620448DEST_PATH_IMAGE014
The score expectation of (a) is,
Figure 803036DEST_PATH_IMAGE016
an intermediate parameter is expressed for measuring the uncertainty of the purchasing strategy pi,
Figure 144019DEST_PATH_IMAGE018
representing a preset weight parameter;
iteratively updating the weight parameters of the purchasing strategy pi, learning to obtain the purchasing strategy pi and a reward function, inputting the state s of the real purchasing short sequence into the deep neural network of the purchasing strategy pi and obtaining a purchasing decisionaWill purchase the decisionaAdding real purchase short sequences to complete length expansion until the length is expanded to a preset length;
updating a purchasing strategy pi by taking the imitation learning as a framework;
based on the commodity pre-embedded representation and the expanded sparse ultra-short user behavior sequence, the information enhancement of the commodity embedded representation is completed by adopting a self-attention model, the user embedded representation is obtained according to the final commodity embedded representation and the user behavior data of the sparse ultra-short user behavior sequence, and personalized recommendation is performed according to the embedded representation of the commodity and the user.
2. The sparse ultrashort sequence-oriented personalized recommendation method as claimed in claim 1, wherein a graph embedding method is adopted to perform embedded representation learning on nodes in the relational graph, and the specific steps include:
constructing a user and commodity relation bipartite graph according to a commodity purchasing history sequence of a user;
generating a commodity purchasing sequence of the user and a commodity purchasing sequence of the user, and constructing an inducing lists x And inducement listss y ；
Dividing a user-commodity relationship bipartite graph into a plurality of induction listss x Ands y the embedded learning is designed as a maximum likelihood problem;
initializing user purchased goods prediction network N x Commodity purchase prediction network N y Attribute-embedded network N a Initializing embedded representation of users or commodities required by an embedding layer, and initializing a weight matrix required by a softmax layer;
from the list of inducerss x Extract an edge from the induction lists x Sampling other edges as the neighbors;
from the list of inducerss y Extract an edge from the list and extract an edge from the lists y Sampling other edges as the neighbors;
setting update target, updating prediction network N x Predictive network N y Attribute-embedded network N a And a commodity or user embedded representation vocabulary;
training until convergence, and updating the weight matrix parameters to obtain the embedded representation of the commodity and the user.
3. The sparse ultrashort sequence-oriented personalized recommendation method as claimed in claim 2, wherein the bipartite graph of the relationship between the user and the commodity is constructed according to the historical sequence of the commodity purchased by the user, and is specifically represented as follows: g (X, Y, A, t), wherein X, Y respectively represents the user node and the commodity node set, A represents the relevant attribute of the user interacting with the commodity, and t represents the time stamp of the interaction.
4. Sparse ultrashort sequence oriented personality according to claim 2The recommendation method is characterized in that the induction list is constructeds x And inducement listss y The method comprises the following specific steps:
generating a sequence x of purchases of goods by a user node x * =[(Y * ,A,t)]And the sequence y of the purchase of the goods y by the user * =[(X * ,A,t)]Wherein X is * To purchase a set of users of Y goods, Y * Obtaining a guidance list for the commodity set purchased by the x user nodes according to the purchase sequence
Figure 871803DEST_PATH_IMAGE020
And inducement lists
Figure 547504DEST_PATH_IMAGE022
Wherein, in the step (A),
Figure 753358DEST_PATH_IMAGE024
is x * And Y * The connection edge is established, and the connection edge is established,
Figure 463694DEST_PATH_IMAGE026
is y and X * And C, establishing a connecting edge, wherein A represents the related attribute of the interaction between the user and the commodity, and t represents the time stamp of the occurrence of the interaction.
5. The sparse ultrashort sequence-oriented personalized recommendation method as claimed in claim 2, wherein the bipartite graph of the relationship between the user and the commodity is divided into a plurality of induction listss x Ands y the embedded learning is designed as a maximum likelihood problem, which is specifically expressed as:
Figure 627959DEST_PATH_IMAGE028
wherein the content of the first and second substances,
Figure 541688DEST_PATH_IMAGE030
is based on experienceBalance parameters set in advance fors x Ands y is to be weighed against the importance of,
Figure 797570DEST_PATH_IMAGE032
are all the parameters that are involved in the model,
Figure 113145DEST_PATH_IMAGE034
representing an objective function;
calculating out
Figure 182732DEST_PATH_IMAGE036
Specifically, it is represented as:
Figure 833025DEST_PATH_IMAGE038
computing
Figure 646260DEST_PATH_IMAGE040
Specifically, it is represented as:
Figure 65609DEST_PATH_IMAGE042
wherein the content of the first and second substances,
Figure 571677DEST_PATH_IMAGE044
and
Figure 459999DEST_PATH_IMAGE046
respectively represent x andy i an embedded representation of a node is shown,
Figure 529455DEST_PATH_IMAGE044
an embedding matrix from the set of user nodes X,
Figure 819622DEST_PATH_IMAGE048
from commodity nodesAn embedding matrix of set Y, the computation of which takes into accounty i Properties of nodesd i ，d i Is an attribute embedded network N a The output result of (a) is obtained,bis the term of the offset, and,
Figure 686471DEST_PATH_IMAGE050
as softmax weighting matrices
Figure 62089DEST_PATH_IMAGE052
The k-th row of the matrix is transposed,
Figure 669657DEST_PATH_IMAGE054
as softmax weighting matrices
Figure 79909DEST_PATH_IMAGE055
First of the matrixlThe lines are transposed so that the lines are transposed,
Figure 380309DEST_PATH_IMAGE057
is taking into account the attributesd j Softmax weight matrix of
Figure 243223DEST_PATH_IMAGE058
The line k of (2) is transposed,
Figure 451220DEST_PATH_IMAGE060
is taking into account the attributesd j Softmax weight matrix of
Figure 184820DEST_PATH_IMAGE062
To (1) alThe lines are transposed so that the lines are transposed,
Figure 653192DEST_PATH_IMAGE064
is taking into account the attributesd j Softmax weight matrix of
Figure 800140DEST_PATH_IMAGE066
To (1) aThe transposition of the k lines is carried out,
Figure 765822DEST_PATH_IMAGE068
is taking into account the attributesd j Softmax weight matrix of
Figure 868776DEST_PATH_IMAGE066
To (1) alThe lines are transposed so that the lines are transposed,
Figure 996132DEST_PATH_IMAGE070
is a preset hyper-parameter, representing the number of neighbors,a i represents x andy i the associated attributes of the interaction of (a),a j represents x andy j x represents a user node,y i indicating merchandiseiThe node(s) of (a) is (are),y j indicating merchandisejA node of (2);
v y andv xi respectively representyAndx i an embedded representation of a node is shown,v y an embedding matrix from the commodity node set Y,
Figure 879643DEST_PATH_IMAGE072
an embedding matrix from the set of user nodes X, the computation of which takes into accountx i Properties of nodesd i ，
Figure 117857DEST_PATH_IMAGE074
As softmax weight matrix
Figure 606476DEST_PATH_IMAGE076
The transpose of the k-th line of (1),
Figure 701471DEST_PATH_IMAGE078
as softmax weight matrix
Figure 26274DEST_PATH_IMAGE080
To (1) alAnd (5) line transposition.
6. The sparse ultrashort sequence-oriented personalized recommendation method of claim 2, wherein the updated weight matrix parameter is expressed as:
Figure 320376DEST_PATH_IMAGE082
wherein the content of the first and second substances,
Figure 148655DEST_PATH_IMAGE084
represents iteration totThe parameters of the model at the time of generation,
Figure 929398DEST_PATH_IMAGE086
is the direction of decrease of the parameter,
Figure 210338DEST_PATH_IMAGE088
the step size is represented as a function of time,
Figure 570781DEST_PATH_IMAGE030
are balance parameters that are preset empirically.
7. The sparse ultrashort sequence-oriented personalized recommendation method as claimed in claim 1, wherein the information enhancement of the commodity embedded representation is completed by adopting a self-attention model, the user embedded representation is obtained according to the final commodity embedded representation and the user behavior data of the sparse ultrashort user behavior sequence, and personalized recommendation is performed according to the embedded representation of the commodity and the user, and the specific steps include:
inputting a real commodity purchasing length sequence of a user and a commodity purchasing length sequence of a virtual user after the short sequence is expanded in a self-attention model;
performing mask preprocessing on the training samples;
training a self-attention model to be convergent, inputting a historical sequence of commodities purchased by each user for the model to obtain an embedded representation of the commodities appearing in the sequence, and aggregating different embedded representations of the same commodity by adopting a mean aggregation or weighted aggregation mode to obtain a final embedded representation of the commodity;
obtaining user embedded representation by adopting a mean value aggregation commodity final embedded representation mode according to a historical commodity purchasing sequence of a real user;
and calculating and sequencing the cosine similarity of the users and the commodities, and recommending the commodity with the highest similarity to each user.
8. A sparse ultrashort sequence-oriented personalized recommendation system is characterized by comprising: the system comprises a user behavior sequence building module, a relational graph building module, an embedded representation learning module, an expert database building module, an imitation learning module, a sequence expansion module, an embedded representation output module and a personalized recommendation module;
the user behavior sequence construction module is used for acquiring sparse ultrashort user behavior data and constructing a sparse ultrashort user behavior sequence;
the relation graph building module is used for building a relation graph of the user and the commodity according to historical data of interaction between the user and the commodity;
the embedded representation learning module is used for embedding representation learning of the nodes in the relational graph by adopting a graph embedding method;
the expert database construction module is used for constructing an expert database based on sparse ultra-short user behavior data;
the imitation learning module is used for learning the purchasing strategy of the expert database by adopting an imitation learning method;
learning the purchasing strategy of the expert database by adopting a simulated learning method, which specifically comprises the following steps:
setting a sequence length threshold, dividing a sparse ultrashort user behavior sequence into a long sequence and a short sequence according to the sequence length threshold, and storing the long sequence into an expert database;
learning the purchasing strategy of the expert database by adopting a simulated learning method based on a generated countermeasure network;
sampling a real purchase sequence state s from a sparse ultra-short user behavior sequence, and obtaining a purchase decision a by utilizing an initialized purchase strategy pi to obtain a generation experience (s, a);
sampling a long sequence from an expert database, segmenting a part of the long sequence containing the first m commodities as a state S, and taking any one of the rest commodities as an A to obtain expert experience (S, A);
the production experience (S, a) and the expert experience (S, a) are simultaneously input into the discriminator D, and the difference is calculated using the cross entropy:
Figure 315883DEST_PATH_IMAGE089
wherein the content of the first and second substances,
Figure 221522DEST_PATH_IMAGE004
the result of the calculation of the difference is represented,
Figure 239025DEST_PATH_IMAGE006
for true purchase sequence statussSelecting actions according to policy
Figure 685050DEST_PATH_IMAGE090
The probability of (a) of (b) being,
Figure 757697DEST_PATH_IMAGE091
for the score expectation of the purchase strategy pi,
Figure 568658DEST_PATH_IMAGE092
to use the strategy
Figure 886507DEST_PATH_IMAGE093
The score expectation of (a) is,
Figure 323173DEST_PATH_IMAGE094
representing intermediate parameters for measuring purchase strategiesThe uncertainty of pi is determined by the number of the pixels,
Figure 980550DEST_PATH_IMAGE095
representing a preset weight parameter;
iteratively updating the weight parameters of the purchasing strategy pi, learning to obtain the purchasing strategy pi and a reward function, inputting the state s of the real purchasing short sequence into the deep neural network of the purchasing strategy pi and obtaining a purchasing decisionaWill purchase the decisionaAdding real purchase short sequences to complete length expansion until the length is expanded to a preset length;
updating a purchasing strategy pi by taking the simulation learning as a framework;
the sequence expansion module is used for completing the expansion of the sparse ultrashort user behavior sequence by adopting a purchase strategy;
the embedded representation output module is used for finishing information enhancement of commodity embedded representation by adopting a self-attention model based on commodity pre-embedded representation and the expanded sparse ultra-short user behavior sequence, and obtaining user embedded representation according to the final commodity embedded representation and the user behavior data of the sparse ultra-short user behavior sequence;
the personalized recommendation module is used for performing personalized recommendation according to the embedded representations of the commodities and the users.
9. A computer-readable storage medium storing a program, wherein the program, when executed by a processor, implements the sparse ultrashort sequence-oriented personalized recommendation method as recited in any one of claims 1 to 7.
10. A computer device comprising a processor and a memory for storing a program executable by the processor, wherein the processor, when executing the program stored in the memory, implements the sparse ultrashort sequence oriented personalized recommendation method as claimed in any one of claims 1 to 7.
CN202210604217.8A 2022-05-31 2022-05-31 Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device Active CN114692005B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210604217.8A CN114692005B (en) 2022-05-31 2022-05-31 Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210604217.8A CN114692005B (en) 2022-05-31 2022-05-31 Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device

Publications (2)

Publication Number Publication Date
CN114692005A CN114692005A (en) 2022-07-01
CN114692005B true CN114692005B (en) 2022-08-12

Family

ID=82131329

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210604217.8A Active CN114692005B (en) 2022-05-31 2022-05-31 Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device

Country Status (1)

Country Link
CN (1) CN114692005B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10019525B1 (en) * 2017-07-26 2018-07-10 International Business Machines Corporation Extractive query-focused multi-document summarization
CN110751557A (en) * 2019-10-10 2020-02-04 中国建设银行股份有限公司 Abnormal fund transaction behavior analysis method and system based on sequence model
CN114202061A (en) * 2021-12-01 2022-03-18 北京航空航天大学 Article recommendation method, electronic device and medium based on generation of confrontation network model and deep reinforcement learning
CN114510653A (en) * 2022-04-21 2022-05-17 中国科学技术大学 Social group recommendation method, system, device and storage medium

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11195057B2 (en) * 2014-03-18 2021-12-07 Z Advanced Computing, Inc. System and method for extremely efficient image and pattern recognition and artificial intelligence platform

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10019525B1 (en) * 2017-07-26 2018-07-10 International Business Machines Corporation Extractive query-focused multi-document summarization
CN110751557A (en) * 2019-10-10 2020-02-04 中国建设银行股份有限公司 Abnormal fund transaction behavior analysis method and system based on sequence model
CN114202061A (en) * 2021-12-01 2022-03-18 北京航空航天大学 Article recommendation method, electronic device and medium based on generation of confrontation network model and deep reinforcement learning
CN114510653A (en) * 2022-04-21 2022-05-17 中国科学技术大学 Social group recommendation method, system, device and storage medium

Also Published As

Publication number Publication date
CN114692005A (en) 2022-07-01

Similar Documents

Publication Publication Date Title
Li et al. A hybrid approach for forecasting ship motion using CNN–GRU–AM and GCWOA
CN108596774B (en) Social information recommendation algorithm and system based on deep network embedded features
Salehinejad et al. Customer shopping pattern prediction: A recurrent neural network approach
CN111695965B (en) Product screening method, system and equipment based on graphic neural network
CN109743196B (en) Network characterization method based on cross-double-layer network random walk
CN111582538B (en) Community value prediction method and system based on graph neural network
KR101926403B1 (en) Promotion Performance Prediction and Recommendation Apparatus in Online Shopping Mall Using Artificial Intelligence
CN110737730B (en) User classification method, device, equipment and storage medium based on unsupervised learning
KR102072820B1 (en) Promotional image improvement apparatus and method in online shopping mall using artificial intelligence
CN113222711B (en) Commodity information recommendation method, system and storage medium
CN114298851A (en) Network user social behavior analysis method and device based on graph sign learning and storage medium
CN115186097A (en) Knowledge graph and reinforcement learning based interactive recommendation method
Wang et al. Optimizing deep belief echo state network with a sensitivity analysis input scaling auto-encoder algorithm
KR102161294B1 (en) Promotion Performance Prediction and Recommendation Apparatus in Online Shopping Mall Using Artificial Intelligence
CN112819024A (en) Model processing method, user data processing method and device and computer equipment
Vargas-Pérez et al. Deep reinforcement learning in agent-based simulations for optimal media planning
Liu Deep learning in marketing: a review and research agenda
Huang et al. Regret theory-based multivariate fusion prediction system and its application to interest rate estimation in multi-scale information systems
CN113569059A (en) Target user identification method and device
CN114692005B (en) Sparse ultrashort sequence-oriented personalized recommendation method, system, medium and device
CN116975686A (en) Method for training student model, behavior prediction method and device
Necula Deep learning for distribution channels’ management
CN116402352A (en) Enterprise risk prediction method and device, electronic equipment and medium
CN116362836A (en) Agricultural product recommendation algorithm based on user behavior sequence
Rawat et al. Role of Mathematics in Novel Artificial Intelligence Realm

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