CN110119467A - A kind of dialogue-based item recommendation method, device, equipment and storage medium - Google Patents
A kind of dialogue-based item recommendation method, device, equipment and storage medium Download PDFInfo
- Publication number
- CN110119467A CN110119467A CN201910398877.3A CN201910398877A CN110119467A CN 110119467 A CN110119467 A CN 110119467A CN 201910398877 A CN201910398877 A CN 201910398877A CN 110119467 A CN110119467 A CN 110119467A
- Authority
- CN
- China
- Prior art keywords
- project
- vector
- session
- attention
- sequence
- 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.)
- Granted
Links
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
- G06F16/951—Indexing; Web crawling techniques
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of dialogue-based item recommendation method, device, equipment and computer readable storage mediums;In this programme, go out an oriented session structure figure by the historical session sequence construct of user, based on the session structure figure, figure neural network can capture the conversion between adjacent items, and generate the hidden state vector of all nodes in the figure, then long-distance dependence is modeled with from attention mechanism, the probability clicked next time is finally predicted in the linear combination of the global preferences of user and current local preference as the implicit vector of current sessions, this programme utilizes the complementary advantage from attention network and figure neural network, realize the accurate recommendation to project.
Description
Technical field
The present invention relates to project recommendation technical field, more specifically to a kind of dialogue-based item recommendation method,
Device, equipment and computer readable storage medium.
Background technique
As development of Mobile Internet technology continues to develop, the information content in network is extended and is increased, net rapidly with exponential law
Information overload and information puzzle problem on network are got worse.In order to provide the user with satisfied information and service, recommender system
It comes into being, becomes the research field of numerous researcher's concerns.Many existing recommender systems are the history based on user
Interaction record, however in numerous application scenarios, user identity be it is unknown, in the session that one carries out, only have here
The historical record of user.In order to solve this problem, dialogue-based recommendation is suggested, and the purpose is to be based on user current
Historical behavior in session predicts the behavior next time of the user, for example clicks the behavior of a project.
Due to the more practical value with height, many dialogue-based recommended methods are suggested.Markov Chain (MC)
It is a typical example, it is assumed that the behavior before behavior is to rely on next time of user.Based on such a strong
It is assumed that user go over behavior the independently combinable accuracy that may will limit recommendation.Nearest research is highlighted based on meeting
The importance of Recognition with Recurrent Neural Network (Recurrent Neural Network, RNN) is used in the recommender system of words, and is obtained
Promising result.For example, Hidasi et al. proposes to use a mutation of RNN, gating cycle unit (Gated
Recurrent Unit, GRU) short-term preference is modeled, then Tan et al. proposes that modified version further increases model originally
Recommendation performance.Li et al. people is also by the sequence pattern for capture user simultaneously with global and local RNN and mainly
Purpose.However, these existing methods usually model the unidirectional conversion between continuous item, have ignored between entire session sequence
Complicated conversion.
Therefore, how more accurately to predict that user clicks the probability of next project, improves the performance of recommendation, is this field
Technical staff's problem to be solved.
Summary of the invention
The purpose of the present invention is to provide a kind of dialogue-based item recommendation method, device, equipment and computer-readable
Storage medium more accurately predicts that user clicks the probability of next project to realize, improves the performance of recommendation.
To achieve the above object, the present invention provides a kind of dialogue-based item recommendation method, comprising:
The historical session sequence of user is obtained, includes each item to sort according to access time in the historical session sequence
Purpose characterizes vector;
Session structure figure corresponding with the historical session sequence is created, and determines each item using the session structure figure
Purpose weights link vector;
By the characterization sequence inputting of the last one project in the weighting link vector and historical session sequence of each project
Figure neural network obtains the hidden state vector for indicating each project of user partial preference;
By the hidden state vector input of each project from attention network, each item for indicating user's global preferences is obtained
Purpose characterizes vector from attention;
Vector is characterized from attention according to each project of hidden state vector sum of each project, determines each project
Session characterizes vector, determines project to be recommended to characterize vector according to the session of each project.
Optionally, creation session structure figure corresponding with the historical session sequence, and utilize the session structure
Figure determines the weighting link vector of each project, comprising:
According to the access order of project each in historical session sequence, oriented session structure figure is determined;
Determine that the input of the session structure figure is weighting connection matrix and connection matrix when exporting;
Using input connection matrix, the historical session sequence when weighting connection matrix, the output, determine each
The weighting link vector of a project.
Optionally, from attention network, obtaining indicates that user is global for the hidden state vector input by each project
Each project of preference characterizes vector from attention, comprising:
The hidden state vector of each project is inputted into multilayer from attention layer;Wherein, each layer is from the defeated of attention layer
It is out the input of next layer adjacent thereto from attention layer;
The last layer is characterized into vector from attention as each project from the output of attention layer.
Optionally, vector is characterized from attention according to each project of hidden state vector sum of each project, determined each
The session of a project characterizes vector, determines project to be recommended to characterize vector according to the session of each project, comprising:
According to each project of hidden state vector sum of each project from attention characterize vector and hidden state to
First weight of amount and the second weight that vector is characterized from attention determine the session characterization vector of each project;
The characterization vector that vector and each project are characterized according to the session of each project, under determining that each project becomes
One access item purpose probability value;
Project to be recommended is determined using the probability value of each project.
To achieve the above object, the present invention further provides a kind of dialogue-based project recommendation devices, comprising:
Module is obtained, when including according to access for obtaining the historical session sequence of user, in the historical session sequence
Between the characterization vector of each project that sorts;
Link vector determining module is weighted, for creating session structure figure corresponding with the historical session sequence, and benefit
The weighting link vector of each project is determined with the session structure figure;
Hidden state vector determining module, for by the weighting link vector of each project and historical session sequence most
The characterization sequence inputting figure neural network of the latter project, obtain indicate user partial preference each project hidden state to
Amount;
Vector determining module is characterized from attention, for inputting the hidden state vector of each project from attention net
Network, obtain each project for indicating user's global preferences characterizes vector from attention;
Project recommendation module, for according to each project of hidden state vector sum of each project from attention characterize to
Amount determines the session characterization vector of each project, determines project to be recommended to characterize vector according to the session of each project.
Optionally, the weighting link vector determining module includes:
Session structure figure determination unit determines oriented for the access order according to project each in historical session sequence
Session structure figure;
Connection matrix determination unit, for determining that the input of the session structure figure connects when weighting connection matrix and output
Connect matrix;
Link vector determination unit is weighted, for connecting square when weighting connection matrix, the output using the input
Battle array, the historical session sequence, determine the weighting link vector of each project.
It is optionally, described to characterize vector determining module from attention, comprising:
Information input unit, for the hidden state vector of each project to be inputted multilayer from attention layer;Wherein, each
Layer from the output of attention layer be the input of next layer adjacent thereto from attention layer;
Vector determination unit is characterized from attention, for the exporting as each project from attention layer by the last layer
Vector is characterized from attention.
Optionally, the project recommendation module, comprising:
Session characterizes vector determination unit, for according to each project of hidden state vector sum of each project from paying attention to
Power characterizes the first weight of vector and hidden state vector and the second weight from attention characterization vector, determines each item
Purpose session characterizes vector;
Probability value determination unit, for characterizing the characterization vector of vector and each project according to the session of each project,
Determine that each project becomes next access item purpose probability value;
Project recommendation unit, for determining project to be recommended using the probability value of each project.
To achieve the above object, the present invention further provides a kind of dialogue-based project recommendation equipment, comprising:
Memory, for storing computer program;
Processor realizes such as above-mentioned dialogue-based project for executing when the computer program is executed by processor
The step of recommended method.
To achieve the above object, the present invention further provides a kind of computer readable storage mediums, described computer-readable
It is stored with computer program on storage medium, is realized when the computer program is executed by processor as above-mentioned dialogue-based
The step of item recommendation method.
By above scheme it is found that the dialogue-based item recommendation method of one kind provided in an embodiment of the present invention, device, setting
Standby and computer readable storage medium;In the present solution, going out an oriented session by the historical session sequence construct of user
Structure chart is based on the session structure figure, and figure neural network can capture the conversion between adjacent items, and generate institute in the figure
There is a hidden state vector of respective nodes, then with long-distance dependence is modeled from attention mechanism, finally the complete of user
The probability clicked next time is predicted in the linear combination of office's preference and current local preference as the implicit vector of current sessions, this
Scheme utilizes the complementary advantage from attention network and figure neural network, realizes the accurate recommendation to project.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
It obtains other drawings based on these drawings.
Fig. 1 is the dialogue-based item recommendation method flow diagram of one kind disclosed by the embodiments of the present invention;
Fig. 2 is session structure figure disclosed by the embodiments of the present invention;
Fig. 3 is that session structure figure disclosed by the embodiments of the present invention inputs side weighting connection matrix schematic diagram;
Fig. 4 is that session structure figure disclosed by the embodiments of the present invention exports side weighting connection matrix schematic diagram;
Fig. 5 is the circuit theory schematic diagram of the dialogue-based project recommendation model of one kind disclosed by the embodiments of the present invention;
Fig. 6 is the dialogue-based project recommendation apparatus structure schematic diagram of one kind disclosed by the embodiments of the present invention;
Fig. 7 is the dialogue-based project recommendation device structure schematic diagram of one kind disclosed by the embodiments of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
Currently, a new series model-Transformer, achieved in various translation duties advanced performance and
Efficiency.Transformer does not use recurrence or convolution operation, but utilizes by stacking the encoder-formed from attention network
Decoder architecture draws the global dependence between outputting and inputting.Self-consciou power is as a kind of special attention machine
System has been widely used for modeling sequence data, and achieves significant achievement, such as machine translation, emotion in numerous applications
Analysis and sequence of recommendation.The success of Transformer model is attributed to the fact that its self-consciou power network, the network pass through weighted average
Operation fully considers all items.It is this to operate the distribution for having dispersed attention although achieving success at present, cause to lack
To the local dependence of adjacent items, and limit the ability that its context for learning project indicates.However the part of adjacent items
Contextual information has been proved to that the dependence between its modeling nerve characterization can be enhanced, especially for attention mould
Type.
Therefore in application, self-consciou power network will be strengthened by figure neural network, propose a figure context certainly
My attention network frame carries out dialogue-based recommendation, and its essence is utilize SAN (Self-Attention Network
(SAN), from attention network) participate in the project of all positions explicitly to capture long-range dependence, utilize figure neural network (Graph
Neural Network, GNN) attributive character on coding nodes and side provides local context information abundant, it is utilized respectively
The advantage of the two networks complementation, the high-order composite character finally learnt in this way can sufficiently reflect user to the inclined of project
It is good, precisely predict that user clicks the probability of next project, to improve the performance of recommendation.
It is understood that the purpose of dialogue-based recommendation is the behavior sequence current based on user, predict under the user
Once will click on which project;Behavior sequence is the historical session sequence in this programme.In the sequence that the present invention considers
Hold the project intersection record of mainly each anonymous session.Provide first | V | a project for being related to all sessions is indexed to project
Set V is represented by V={ v1,v2,...,v|V|, wherein V indicates the set of all items, and v1, v2 represent each
Mesh.Project in this programme can refer to any recommended article, such as commodity, music etc..For each anonymous session,
The behavior sequence that user clicks according to time sequence is represented by S={ s1,s2,...,sn, wherein st∈ V indicate the user when
Between t click project.
According to as defined above, the problem of present invention research, can sum up are as follows: for the S for the behavior sequence being truncated in time tt
={ s1,s2,...,st-1,st(1≤t < n), project (the i.e. s that prediction may be clicked next timet+1).Specifically, this programme
The ranked list occurred in all candidate items of current sessions can be generated.Indicate all items
Output probability, whereinCorresponding to project viRecommender score.Because nominator is usually that user proposes multiple recommendations,
This project can be fromMiddle top-N project of selection is recommended.
Referring to Fig. 1, the dialogue-based item recommendation method of one kind provided in an embodiment of the present invention, comprising:
S101, the historical session sequence for obtaining user, include sorting according to access time in the historical session sequence
The characterization vector of each project;
It should be noted that the historical session sequence in this programme is above: S={ s1,s2,...,sn,
In, the s that includes in sequence1,s2,...,snIt is sorted according to access time, that is to say, that s1For the project accessed at first
Characterization vector.
S102, creation session structure figure corresponding with the historical session sequence, and determined using the session structure figure
The weighting link vector of each project;
Wherein, creation session structure figure corresponding with the historical session sequence, and utilize the session structure figure
Determine the weighting link vector of each project, comprising:
According to the access order of project each in historical session sequence, oriented session structure figure is determined;
Determine that the input of the session structure figure is weighting connection matrix and connection matrix when exporting;
Using input connection matrix, the historical session sequence when weighting connection matrix, the output, determine each
The weighting link vector of a project.
It is understood that figure neural network GNN can be operated in digraph, study graph structure back end
Characterization captures the contextual information of sequence part.Therefore, the present invention captures user partial dependence using GNN model.It is transporting
Before GNN, need to construct a significant figure to all session datas.Give a session S={ s1,s2,...,
sn, by each project siIt is expressed as a node, (si-1,si) it is expressed as a line in figure, it is meant that a user is in session
Project s has been clicked in Si-1Project s is clicked lateri.Therefore, each session can be modeled as a digraph, the digraph
For session structure figure, graph structure then is updated by promoting the communication between different nodes.
Specifically, M is allowedI,MO∈Rn×nIt respectively indicates input in session diagram and connects square when weighting connection matrix with output
Battle array, wherein R is set of real numbers, and n indicates the quantity of project.For example, it is assumed that a session S={ s1,s3,s2,s4,s3, referring to fig. 2,
Side is inputted for session structure figure disclosed by the embodiments of the present invention and is added referring to Fig. 3 for session structure figure disclosed by the embodiments of the present invention
Connection matrix schematic diagram is weighed, referring to fig. 4, side is exported for session structure figure disclosed by the embodiments of the present invention and weights connection matrix signal
Figure;As can be seen that since several projects may be repetitively appearing in some session sequence, this programme is each edge
Allocation criterion weight, the weight by this while start node out-degree divided by while total frequency of occurrence obtained by.It needs to illustrate
It is that the model for the progress project recommendation that this programme provides can support the various strategies of construction session diagram, and generate and connect accordingly
Matrix is connect, is not limited specifically in this programme.
It should be noted that initially each project can give one one-dimensional vector of its random initializtion, length is d, indicates this
The low-dimensional of a project implies vector and each project v ∈ V is further transformed into the hidden vector space of unified low-dimensional, converts
Afterwards, pass through knot vector s ∈ RdThe d dimension real value of expression project v implies vector;The knot vector is the item in this programme
Purpose characterizes vector.
Session structure figure is constructed, input side is obtained and weights connection matrix MIWith output side connection matrix MOAfterwards, it needs according to defeated
Enter the weighting link vector S={ s of connection matrix and each project when weighting connection matrix, output1,s2,...,snCalculate respectively
The weighting link vector a of a projectt;For the node s under the time t in session diagramt, the information propagation between different nodes can
To be formulated as:
at=Concat (Mt I([s1,...,sn]Wa I+bI),
Mt O([s1,...,sn]Wa O+bO). (1)
Wherein, two number groups are stitched together by Concat expression, such as [1,2,3] and [4,5,6], are exactly after Concat
[1,2,3,4,5,6];Wa I,Wa O∈Rd×dIt is the first parameter matrix and the second parameter matrix respectively,
bI,bO∈RdIt is the first bias vector and the second bias vector respectively.Mt I,Mt O∈R1×nCorrespond to node st's
Input the t row of the connection matrix in the t row and output for weighting connection matrix.atIt is extracted node stAbove and below neighbor node
Literary information.Then by atWith previous state knot vector st-1It inputs and obtains a new project hidden state into figure neural network
Vector ht, modeled the local preference of user;In the present solution, for a session sequence: s1,s2,...,sn, as calculating s2
When s1 be exactly previous state node.
S103, the characterization sequence for weighting the last one project in link vector and historical session sequence by each project
Figure neural network is inputted, the hidden state vector for indicating each project of user partial preference is obtained;
Specifically, this programme mainly obtains the implicit vector of each node by GNN figure neural network, gives one
The upper knot vector s of nodet-1With the weighting link vector a of the nodet, the hidden state vector h of final GNN layers of outputt
Are as follows:
zt=σ (Wzat+pzst-1),
rt=σ (Wrat+prst-1),
Wherein, Wz,Wr,Wh∈R2d×dFor the parameter for needing to learn after training, weight is indicated, it can be with first to be learned
Parameter, the second parameter to be learned, third parameter to be learned indicate that d is the dimension of implicit vector, and every dimension represents the project and exists
The value of a certain feature, this be characterized in it is abstract, model oneself study.Pz,Pr,Ph∈Rd×dIt is also weight, is to be learned
Parameter can be indicated with the 4th parameter to be learned, the 5th parameter to be learned, the 6th parameter to be learned.σ () expression is one
Sigmoid layers, the value between 0 to 1 is mapped, 1 represents " being fully retained ", and 0 represents " abandoning completely ",It indicates between two vectors
Hadamard product calculation.zt,rtIt is to update door and resetting door respectively, determines which information retained and abandon.It indicates to hide shape
State, htIt is comparable to out gate.
S104, the hidden state vector of each project is inputted from attention network, obtaining indicates user's global preferences
Each project characterizes vector from attention;
Wherein, from attention network, obtaining indicates that user is global partially for the hidden state vector input by each project
Good each project characterizes vector from attention, comprising:
The hidden state vector of each project is inputted into multilayer from attention layer;Wherein, each layer is from the defeated of attention layer
It is out the input of next layer adjacent thereto from attention layer;
The last layer is characterized into vector from attention as each project from the output of attention layer.
It should be noted that being upper one layer GNN layers of output from the input of attention network.In the present solution, from paying attention to
Power layer is made of three parts: from attention layer, feedforward network and multilayer are from attention.
1, from attention layer:
From attention mechanism it is a special case of attention mechanism, and is successfully applied to every field, example
Such as natural language processing and question answering system.Global dependence pass can be drawn between sequence outputting and inputting from attention mechanism
System, and project-project interaction between capturing sequence itself, but regardless of they distance how.Therefore the present invention uses SAN mould
Type captures user's lasting arrangement dependence.The structure block of self-consciou power network is a scalable dot product attention mechanism.
The input of the attention model is formed by inquiring (Queries), key (Keys) and value (Values).Output be by
Obtained by the weighted sum of the value of the similitude of inquiry and key.In our example, all these three components are all identical, and
It is made of the history intersection record of user.Give an input vector St u, three are converted by input using Linear Mapping first
Matrix, respectively inquiry matrix AQ, key matrix AKWith value matrix AV;Such as: AQ=WQSt u.The output from attention block by with
Lower formula calculates:
Wherein, ATT indicates that a simplified expression, A will be on the right of equationQ,AK,AV∈Rn×dIt can be appreciated that mapping square
Battle array, mapping can make model more flexible.Scale factorIt is to avoid the excessive presentation extremum of inner product value, especially works as dimension
It is played regulatory role when spending high.
Therefore, in this application, obtain in session diagram after the implicit vector of each item nodes, that is, determine H=[h1,
h2,...,hn] after, it needs to be inputted into SAN and preferably captures global session preference, in the present embodiment, the will be inputted
One layer of vector exported from after attention layer indicate with F, namely:
F=ATT (HWQ',HWK',HWV') (4)
Wherein, WQ',WK',WV'∈R2d×dIt can be expressed as the first mapping matrix, the second mapping matrix and third mapping
Matrix, HWQ'It is equivalent to above from the A in attention networkQ。
2, feedforward network:
After one layer of SAN, we assign mould with two layers of linear transformation with ReLu activation primitive
Type is non-linear, and considers the interaction between different implicit dimensions.However, transmission damage may occur in the operation of self-consciou power
It loses.Therefore, we are added to residual error connection after feedforward network, and model is made to be easier to utilize low level information.
E=ReLU (FW1+b1)W2+b2+ F, (5)
Wherein, W1And W2The first weighting matrix and the second weighting matrix of respectively d × d, b1 and b2 be d dimension first partially
See vector sum the second prejudice vector;It should be noted that in the present embodiment, all parameters containing W and b, in spite of having
Subscript is all random initializtion, and then model is trained.In addition, in order to alleviate the over-fitting in feedforward network
Problem has used Dropout technology in training.Here, property for simplicity, entire it is from attention net definitions by above
E=SAN (H) (6)
3, multilayer SAN:
By first layer from after attention layer, F has substantially polymerize the embedding of all prior projects using adaptive weighting
Enter.However, there is work to show different to capture different types of feature from attention layer recently.Therefore, by being based on F
Another self-consciou power layer come learn higher level project conversion may be useful.Therefore, in the present solution, using
With multilayer from attention layer from attention network, wherein kth (k > 1) layer is defined as:
E(k)=SAN (E(k-1)) (7)
Therein, it can be seen that first layer is defined as E(1)=E=SAN (H), each layer later, by upper one layer
Export the input as this layer, E(k)∈Rn×dIt is then the final output of multilayer self-consciou power network.
S105, vector is characterized from attention according to each project of hidden state vector sum of each project, determined each
The session of project characterizes vector, determines project to be recommended to characterize vector according to the session of each project.
Wherein, vector is characterized from attention according to each project of hidden state vector sum of each project, determined each
The session of project characterizes vector, determines project to be recommended to characterize vector according to the session of each project, comprising:
According to each project of hidden state vector sum of each project from attention characterize vector and hidden state to
First weight of amount and the second weight that vector is characterized from attention determine the session characterization vector of each project;
The characterization vector that vector and each project are characterized according to the session of each project, under determining that each project becomes
One access item purpose probability value;
Project to be recommended is determined using the probability value of each project.
It should be noted that by obtaining after extracting the sequence information of session from attention network self-adapting
The long-term of sequence characterizes vector E from attention(k).In order to which the project of user clicked is better anticipated next time, by long-term preference
E(k)With current session interest hnWeighted sum, then the implicit vector using this after combining-session characterization vector is as the session
Expression.For session S={ s1,s2,...,sn, take the output E of multilayer SAN(k)It is last one-dimensional as global insertion vector,
Locally embedding vector is expressed as the implicit vector that closest approach hits project, i.e. hn, then weighted sum is characterized as final session
Vector.
Sf=ω En (k)+(1-ω)hn (8)
Wherein, En (k)∈RdIndicate the line n from attention characterization vector, (1- ω) is the first weight, and ω is the second power
Weight, ω can be preset according to the actual situation between [0,1].Then each candidate items v is predictedi∈ V belongs to next
The probability of secondary click, the probability are to become next access item purpose probability value in this programme:
Wherein,Expression project viAs the recommendation probability clicked in session S next time;The determine the probability of each project
Afterwards, can therefrom select probability value it is maximum N number of as project to be recommended, the value of the N can be 1, or more
A, this programme does not limit specifically, can be adjusted according to the actual situation.It should be noted that this programme obtain finally to
After the project of recommendation, need that cross entropy loss function is used to optimize for model, by minimizing following objective function:
Wherein, y indicates the one-hot coding vector of actual items, and θ is the set of all hyper parameters in model, and λ is canonical
Change weight;To a list entries [s1,s2,...,st,st+1,...,sn], calculating stWhen, y is st+1, calculate s1's
When y be exactly s2。
It in summary it can be seen, for being able to achieve for the model of above-mentioned item recommendation method, be divided into three parts: being respectively
The figure neural net layer part for executing S101-S103, execute S104 from attention layer part, and execute the prediction interval of S105
Part.It is the circuit theory schematic diagram of the dialogue-based project recommendation model of one kind provided in an embodiment of the present invention referring to Fig. 5;
The leftmost side of the frame diagram is the building of user conversation figure, and middle section is the utilization from attention network, and the rightmost side is prediction
Layer, calculates the rank score of all candidate items.
In the present embodiment, two kinds of algorithms to the model training in this programme are provided, respectively gradient descent algorithm and
BPTT (Back-Propagation Through Time)-back-propagation algorithm at any time, is introduced separately below:
(1) gradient descent algorithm:
Gradient descent method is more commonly used first-order optimization method.Gradient descent method is based on the idea that be: finding certain function
Minimum value, the best way is sought along the negative gradient direction of the function.If gradient is denoted as ▽, function f's (x, y)
Gradient may be expressed as:
So the negative gradient opposite direction that means that gradient, this gradient mean to move along the direction of xIt moves in direction along yWherein, function f (x, y) must be defined on point to be calculated and can
It is micro-.
In each iteration, gradient descent algorithm can all be moved along negative gradient direction and be moved a step.Wherein, gradient operator always refers to
Decline most fast direction to functional value.Mentioned here is moving direction, without referring to amount of movement size.The magnitude is known as walking
It is long, it is denoted as α.The iterative formula of gradient decline can indicate are as follows:
w←w-α▽wf(w)
The formula will be iterated always execution, until reaching some stop condition.For example the number of iterations reaches some
Designated value or algorithm reach some error range that can permit.
(2) BPTT- back-propagation algorithm at any time
Most common optimization algorithm is back-propagation algorithm at any time in neural network, is that one kind is suitable for multilayer neuron
A kind of learning algorithm of network.Similar with back-propagation algorithm (BP), it is the gradient descent algorithm of anti-pass in time, is established
On the basis of gradient descent method.The input/output relation of BPTT network is substantially a kind of mapping relations: an input m output
The function completed of BPTT neural network be that the Continuous Mappings of Euclidean space one finite field into m dimension Euclidean space are tieed up from n, this
One mapping has nonlinearity.Its information processing capability is multiple compound from simple non-linear functions, therefore has
Very strong function reappears ability.This is the basis that BPTT algorithm is able to apply.Back-propagation algorithm is mainly by the (excitation of two links
Propagate, weight updates) iterative cycles iteration, until the response to input of network reaches scheduled target zone.
The learning process of BPTT algorithm is made of forward-propagating process and back-propagation process.During forward-propagating,
Information is inputted by input layer through hidden layer, successively handles and is transmitted to output layer.If cannot get desired output in output layer
Value then takes the quadratic sum of output and desired error to be transferred to backpropagation as objective function, successively find out objective function to each
The partial derivative of neuron weight constitutes objective function and measures to the ladder of weight vector, as the foundation of modification weight, the study of network
It is completed during weight modification.When error reaches desired value, e-learning terminates.
In order to facilitate this programme is understood, the concrete application process an of this programme is provided herein:
Session recommendation is briefly exactly that a user accesses a website, but user is not logged in this programme, only
It is the several commodity arbitrarily browsed in webpage, then just has left, at this moment, by Continuous behavior of the user within this short time
Just it is called session.When training the model in this programme, it is necessary to input the session sequence of this anonymity of many items.
Firstly, obtain many session sequences, every session sequence by several item designs, such as commodity, music etc.,
It is referred to as items;Every session sequence is according to time-sequencing.By this all session sequence structure at a session structure
Figure.Then for each individual session [s1,s2,...,sn], it is exactly a subgraph in the session diagram, i.e., the left side Fig. 5 is empty
That a part of line, available two connection matrix MIAnd MO, a then is calculated to the node of each time point tt, will
atAnd st-1It is input in GNN together and obtains ht;Then htIt is input to and obtains E from attention network(k), finally will be from attention
The output E of power layer(k)With the output h of GNNtDo a linear weighted combination, so that it may final output is obtained, as current meeting
The characterization vector S of wordsf, finally predict that current session vector S can be characterized when next behaviorfWith the table of project
Levy vector viMultiplication obtains a scoreSo as to obtain the score of all items, score is highest just as current meeting
Which item the behavior next time of words, i.e. anonymous may click next time.
It in summary it can be seen, the dialogue-based project recommendation scheme of this programme, it is intended to predict to use according to anonymous session
The next step action at family is a key task of many online services, such as e-commerce and Streaming Media.Recently, self is infused
Meaning power network (SAN) due to without using circulation or convolution operation all achieved in terms of various Series Modeling tasks it is great at
Function.But lack local dependence present on adjacent items from attention network, and limit it and learn project in sequence
Context indicate ability.
And how this programme is for using figure neural network and self-consciou power mechanism enhancing dialogue-based performance of recommending
Problem proposes a kind of completely new figure context self-consciou power session recommendation frame.Specifically, for all session sequences
Column, dynamic construction goes out a digraph structure first for we, and based on the session diagram constructed, figure neural network can capture adjacent
Relationship between project, and the hidden vector of corresponding node is generated, following each session passes through new hidden vector and uses certainly
My attention mechanism relies on to learn length of run, and then each session is represented as the linear of the session global preferences and current interest
In conjunction with.Finally, we have evaluated proposed model using the real data set of two different application scenes.
It should be noted that this programme has the following beneficial effects:
1) in order to promote the characterization of session sequence, this programme is based on figure neural network and proposes a novel figure context
From attention model, recommendation performance is improved using from the complementary advantage of attention network and figure neural network.
2) figure neural network is used to model the Local map Structure Dependence of each independent session, and multilayer is from attention network
Context can be obtained locally to characterize.Also, it is tested by two real data sets, to assess mould described in this programme
Type performance, the experimental results showed that, this dialogue-based project recommendation mode that this programme proposes persistently is better than presently relevant
State-of-the-art method.
Project recommendation device provided in an embodiment of the present invention is introduced below, project recommendation device described below with
Above-described item recommendation method can be cross-referenced.
Referring to Fig. 6, the dialogue-based project recommendation device of one kind provided in an embodiment of the present invention, comprising:
It includes according to access in the historical session sequence that module 100, which is obtained, for obtaining the historical session sequence of user
The characterization vector of each project of time-sequencing;
Link vector determining module 200 is weighted, for creating session structure figure corresponding with the historical session sequence, and
The weighting link vector of each project is determined using the session structure figure;
Hidden state vector determining module 300, for will be in the weighting link vector of each project and historical session sequence
The last one project characterization sequence inputting figure neural network, obtain indicate user partial preference each project implicit shape
State vector;
Vector determining module 400 is characterized from attention, for inputting the hidden state vector of each project from attention
Network, obtain each project for indicating user's global preferences characterizes vector from attention;
Project recommendation module 500, for according to each project of hidden state vector sum of each project from attention table
Vector is levied, the session characterization vector of each project is determined, determines item to be recommended to characterize vector according to the session of each project
Mesh.
Wherein, the weighting link vector determining module includes:
Session structure figure determination unit determines oriented for the access order according to project each in historical session sequence
Session structure figure;
Connection matrix determination unit, for determining that the input of the session structure figure connects when weighting connection matrix and output
Connect matrix;
Link vector determination unit is weighted, for connecting square when weighting connection matrix, the output using the input
Battle array, the historical session sequence, determine the weighting link vector of each project.
It is wherein, described to characterize vector determining module from attention, comprising:
Information input unit, for the hidden state vector of each project to be inputted multilayer from attention layer;Wherein, each
Layer from the output of attention layer be the input of next layer adjacent thereto from attention layer;
Vector determination unit is characterized from attention, for the exporting as each project from attention layer by the last layer
Vector is characterized from attention.
Wherein, the project recommendation module, comprising:
Session characterizes vector determination unit, for according to each project of hidden state vector sum of each project from paying attention to
Power characterizes the first weight of vector and hidden state vector and the second weight from attention characterization vector, determines each item
Purpose session characterizes vector;
Probability value determination unit, for characterizing the characterization vector of vector and each project according to the session of each project,
Determine that each project becomes next access item purpose probability value;
Project recommendation unit, for determining project to be recommended using the probability value of each project.
Referring to Fig. 7, the embodiment of the invention also discloses a kind of dialogue-based project recommendation equipment 1, comprising:
Memory 11, for storing computer program;
Processor 12 is realized dialogue-based as described in above method embodiment when for executing the computer program
The step of item recommendation method.
In the present embodiment, equipment 1 can be PC (Personal Computer, PC), be also possible to intelligent hand
The terminal devices such as machine, tablet computer, palm PC, portable computer.
The equipment 1 may include memory 11, processor 12 and bus 13.
Wherein, memory 11 include at least a type of readable storage medium storing program for executing, the readable storage medium storing program for executing include flash memory,
Hard disk, multimedia card, card-type memory (for example, SD or DX memory etc.), magnetic storage, disk, CD etc..Memory 11
It can be the internal storage unit of equipment 1, such as the hard disk of the equipment 1 in some embodiments.Memory 11 is in other realities
Apply the plug-in type hard disk being equipped on the External memory equipment for being also possible to equipment 1 in example, such as equipment 1, intelligent memory card
(Smart Media Card, SMC), secure digital (Secure Digital, SD) card, flash card (Flash Card) etc..Into
One step, memory 11 can also both internal storage units including equipment 1 or including External memory equipment.Memory 11 is not only
It can be used for storing and be installed on application software and Various types of data, such as the code of project recommendation program of equipment 1 etc., can also use
In temporarily storing the data that has exported or will export.
Processor 12 can be in some embodiments a central processing unit (Central Processing Unit,
CPU), controller, microcontroller, microprocessor or other data processing chips, the program for being stored in run memory 11
Code or processing data, such as project implementation recommended program etc..
The bus 13 can be Peripheral Component Interconnect standard (peripheral component interconnect, abbreviation
PCI) bus or expanding the industrial standard structure (extended industry standard architecture, abbreviation EISA)
Bus etc..The bus can be divided into address bus, data/address bus, control bus etc..For convenient for indicating, in Fig. 7 only with one slightly
Line indicates, it is not intended that an only bus or a type of bus.
Further, equipment can also include network interface 14, network interface 14 optionally may include wireline interface and/
Or wireless interface (such as WI-FI interface, blue tooth interface), it is logical commonly used in being established between the equipment 1 and other electronic equipments
Letter connection.
Optionally, which can also include user interface, and user interface may include display (Display), input
Unit such as keyboard (Keyboard), optional user interface can also include standard wireline interface and wireless interface.It is optional
Ground, in some embodiments, display can be light-emitting diode display, liquid crystal display, touch-control liquid crystal display and OLED
(Organic Light-Emitting Diode, Organic Light Emitting Diode) touches device etc..Wherein, display can also be appropriate
Referred to as display screen or display unit, for showing the information handled in the device 1 and for showing visual user interface.
Fig. 7 illustrates only the equipment 1 with component 11-14 and project recommendation program, and those skilled in the art can manage
Solution, the structure shown in Fig. 7 do not constitute the restriction to equipment 1, may include less than diagram or more components, or
Person combines certain components or different component layouts.
The embodiment of the invention also discloses a kind of computer readable storage medium, deposited on the computer readable storage medium
Computer program is contained, is realized when the computer program is executed by processor dialogue-based as described in above method embodiment
Item recommendation method the step of.
Wherein, the storage medium may include: USB flash disk, mobile hard disk, read-only memory (Read-Only Memory,
ROM), random access memory (Random Access Memory, RAM), magnetic or disk etc. are various can store program
The medium of code.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments without departing from the spirit or scope of the present invention.Therefore, of the invention
It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one
The widest scope of cause.
Claims (10)
1. a kind of dialogue-based item recommendation method characterized by comprising
The historical session sequence of user is obtained, includes each project to sort according to access time in the historical session sequence
Characterize vector;
Session structure figure corresponding with the historical session sequence is created, and determines each project using the session structure figure
Weight link vector;
By the characterization sequence inputting figure mind of the last one project in the weighting link vector and historical session sequence of each project
Through network, the hidden state vector for indicating each project of user partial preference is obtained;
By the hidden state vector input of each project from attention network, each project for indicating user's global preferences is obtained
Vector is characterized from attention;
Vector is characterized from attention according to each project of hidden state vector sum of each project, determines the session of each project
Vector is characterized, determines project to be recommended to characterize vector according to the session of each project.
2. item recommendation method according to claim 1, which is characterized in that the creation and the historical session sequence pair
The session structure figure answered, and determine using the session structure figure weighting link vector of each project, comprising:
According to the access order of project each in historical session sequence, oriented session structure figure is determined;
Determine that the input of the session structure figure is weighting connection matrix and connection matrix when exporting;
Using input connection matrix, the historical session sequence when weighting connection matrix, the output, each item is determined
Purpose weights link vector.
3. item recommendation method according to claim 2, which is characterized in that the hidden state vector by each project
From attention network, obtain each project for indicating user's global preferences characterizes vector from attention for input, comprising:
The hidden state vector of each project is inputted into multilayer from attention layer;Wherein, the output of each layer from attention layer is
The input of next layer adjacent thereto from attention layer;
The last layer is characterized into vector from attention as each project from the output of attention layer.
4. item recommendation method as claimed in any of claims 1 to 3, which is characterized in that according to each project
The each project of hidden state vector sum characterizes vector from attention, the session characterization vector of each project is determined, according to each
The session characterization vector of a project determines project to be recommended, comprising:
Vector and hidden state vector are characterized from attention according to each project of hidden state vector sum of each project
First weight and the second weight that vector is characterized from attention determine the session characterization vector of each project;
The characterization vector that vector and each project are characterized according to the session of each project determines that each project becomes next
Access item purpose probability value;
Project to be recommended is determined using the probability value of each project.
5. a kind of dialogue-based project recommendation device characterized by comprising
It includes arranging according to access time in the historical session sequence that module, which is obtained, for obtaining the historical session sequence of user
The characterization vector of each project of sequence;
Link vector determining module is weighted, for creating session structure figure corresponding with the historical session sequence, and utilizes institute
State the weighting link vector that session structure figure determines each project;
Hidden state vector determining module, for last in the weighting link vector and historical session sequence by each project
The characterization sequence inputting figure neural network of a project obtains the hidden state vector for indicating each project of user partial preference;
Vector determining module being characterized from attention, for inputting the hidden state vector of each project from attention network, being obtained
Must indicate each project of user's global preferences characterizes vector from attention;
Project recommendation module, for characterizing vector from attention according to each project of hidden state vector sum of each project,
The session characterization vector for determining each project, determines project to be recommended to characterize vector according to the session of each project.
6. project recommendation device according to claim 5, which is characterized in that the weighting link vector determining module packet
It includes:
Session structure figure determination unit determines oriented meeting for the access order according to project each in historical session sequence
Talk about structure chart;
Connection matrix determination unit, for determining that the input of the session structure figure connects square when weighting connection matrix with output
Battle array;
Link vector determination unit is weighted, for utilizing input connection matrix, institute when weighting connection matrix, the output
Historical session sequence is stated, determines the weighting link vector of each project.
7. project recommendation device according to claim 6, which is characterized in that described to determine mould from attention characterization vector
Block, comprising:
Information input unit, for the hidden state vector of each project to be inputted multilayer from attention layer;Wherein, each layer from
The output of attention layer is the input of next layer adjacent thereto from attention layer;
Vector determination unit is characterized from attention, for output the infusing certainly as each project by the last layer from attention layer
Power of anticipating characterizes vector.
8. the project recommendation device according to any one of claim 5 to 7, which is characterized in that the project recommendation mould
Block, comprising:
Session characterizes vector determination unit, for according to each project of hidden state vector sum of each project from attention table
The first weight of vector and hidden state vector and the second weight from attention characterization vector are levied, determines each project
Session characterizes vector;
Probability value determination unit is determined for characterizing the characterization vector of vector and each project according to the session of each project
Each project becomes next access item purpose probability value;
Project recommendation unit, for determining project to be recommended using the probability value of each project.
9. a kind of dialogue-based project recommendation equipment characterized by comprising
Memory, for storing computer program;
Processor realizes such as Claims 1-4 described in any item dialogue-based items when for executing the computer program
The step of mesh recommended method.
10. a kind of computer readable storage medium, which is characterized in that be stored with computer on the computer readable storage medium
Program realizes such as Claims 1-4 described in any item dialogue-based projects when the computer program is executed by processor
The step of recommended method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910398877.3A CN110119467B (en) | 2019-05-14 | 2019-05-14 | Project recommendation method, device, equipment and storage medium based on session |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910398877.3A CN110119467B (en) | 2019-05-14 | 2019-05-14 | Project recommendation method, device, equipment and storage medium based on session |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110119467A true CN110119467A (en) | 2019-08-13 |
CN110119467B CN110119467B (en) | 2023-05-02 |
Family
ID=67522374
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910398877.3A Active CN110119467B (en) | 2019-05-14 | 2019-05-14 | Project recommendation method, device, equipment and storage medium based on session |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110119467B (en) |
Cited By (38)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110598118A (en) * | 2019-09-23 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Resource object recommendation method and device and computer readable medium |
CN110619082A (en) * | 2019-09-20 | 2019-12-27 | 苏州市职业大学 | Project recommendation method based on repeated search mechanism |
CN110633789A (en) * | 2019-08-27 | 2019-12-31 | 苏州市职业大学 | Self-attention network information processing method for streaming media recommendation |
CN110705709A (en) * | 2019-10-14 | 2020-01-17 | 支付宝(杭州)信息技术有限公司 | Method and device for training neural network model of graph |
CN110717098A (en) * | 2019-09-20 | 2020-01-21 | 中国科学院自动化研究所 | Meta-path-based context-aware user modeling method and sequence recommendation method |
CN110795618A (en) * | 2019-09-12 | 2020-02-14 | 腾讯科技(深圳)有限公司 | Content recommendation method, device, equipment and computer readable storage medium |
CN110796313A (en) * | 2019-11-01 | 2020-02-14 | 北京理工大学 | Session recommendation method based on weighted graph volume and item attraction model |
CN110807156A (en) * | 2019-10-23 | 2020-02-18 | 山东师范大学 | Interest recommendation method and system based on user sequence click behaviors |
CN110826700A (en) * | 2019-11-13 | 2020-02-21 | 中国科学技术大学 | Method for realizing and classifying bilinear graph neural network model for modeling neighbor interaction |
CN111046257A (en) * | 2019-12-09 | 2020-04-21 | 北京百度网讯科技有限公司 | Session recommendation method and device and electronic equipment |
CN111179031A (en) * | 2019-12-23 | 2020-05-19 | 第四范式(北京)技术有限公司 | Training method, device and system for commodity recommendation model |
CN111242748A (en) * | 2020-02-21 | 2020-06-05 | 腾讯科技(深圳)有限公司 | Method, apparatus, and storage medium for recommending items to a user |
CN111259256A (en) * | 2020-02-13 | 2020-06-09 | 腾讯科技(深圳)有限公司 | Content processing method, content processing device, computer-readable storage medium and computer equipment |
CN111460331A (en) * | 2020-04-07 | 2020-07-28 | 中国计量大学 | Conversation recommendation method based on space-time diagram |
CN111461301A (en) * | 2020-03-30 | 2020-07-28 | 北京沃东天骏信息技术有限公司 | Serialized data processing method and device, and text processing method and device |
CN111538906A (en) * | 2020-05-29 | 2020-08-14 | 支付宝(杭州)信息技术有限公司 | Information pushing method and device based on privacy protection |
CN111667067A (en) * | 2020-05-28 | 2020-09-15 | 平安医疗健康管理股份有限公司 | Recommendation method and device based on graph neural network and computer equipment |
CN111708876A (en) * | 2020-06-16 | 2020-09-25 | 北京百度网讯科技有限公司 | Method and device for generating information |
CN111737578A (en) * | 2020-06-22 | 2020-10-02 | 陕西师范大学 | Recommendation method and system |
CN111813640A (en) * | 2020-03-25 | 2020-10-23 | 北京嘀嘀无限科技发展有限公司 | Method, apparatus, electronic device and storage medium for predicting user behavior |
CN111859160A (en) * | 2020-08-07 | 2020-10-30 | 成都理工大学 | Method and system for recommending session sequence based on graph neural network |
CN111949865A (en) * | 2020-08-10 | 2020-11-17 | 杭州电子科技大学 | Interest point recommendation method based on graph neural network and user long-term and short-term preference |
CN112084404A (en) * | 2020-09-01 | 2020-12-15 | 北京百度网讯科技有限公司 | Content recommendation method, device, equipment and medium |
CN112150210A (en) * | 2020-06-19 | 2020-12-29 | 南京理工大学 | Improved neural network recommendation method and system based on GGNN (global warming network) |
CN112256879A (en) * | 2020-10-29 | 2021-01-22 | 贝壳技术有限公司 | Information processing method and apparatus, electronic device, and computer-readable storage medium |
CN112258262A (en) * | 2020-09-15 | 2021-01-22 | 浙江大学 | Conversation recommendation method based on convolution self-attention network |
WO2021038592A2 (en) | 2019-08-30 | 2021-03-04 | Tata Consultancy Services Limited | System and method for handling popularity bias in item recommendations |
CN112541639A (en) * | 2020-12-22 | 2021-03-23 | 宜宾电子科技大学研究院 | Recommendation system scoring prediction method based on graph neural network and attention mechanism |
CN112559878A (en) * | 2020-12-24 | 2021-03-26 | 山西大学 | Sequence recommendation system and recommendation method based on graph neural network |
CN112579914A (en) * | 2020-12-31 | 2021-03-30 | 安徽大学 | Time perception service recommendation system and method based on self-attention factorization machine |
CN112685657A (en) * | 2021-01-06 | 2021-04-20 | 中国计量大学 | Conversation social recommendation method based on multi-mode cross fusion graph network |
CN112733018A (en) * | 2020-12-31 | 2021-04-30 | 哈尔滨工程大学 | Session recommendation method based on graph neural network GNN and multi-task learning |
CN112765461A (en) * | 2021-01-12 | 2021-05-07 | 中国计量大学 | Session recommendation method based on multi-interest capsule network |
CN112966193A (en) * | 2021-03-05 | 2021-06-15 | 北京百度网讯科技有限公司 | Travel intention inference method, model training method, related device and electronic equipment |
CN113467740A (en) * | 2021-07-15 | 2021-10-01 | 青岛博天数通信息科技有限公司 | Video monitoring array display optimization method and device based on joint coding |
CN114743053A (en) * | 2022-04-14 | 2022-07-12 | 电子科技大学 | Magnetic resonance image auxiliary processing system based on graph neural network and self attention |
WO2023108324A1 (en) * | 2021-12-13 | 2023-06-22 | 中国科学院深圳先进技术研究院 | Comparative learning enhanced two-stream model recommendation system and algorithm |
CN116521972A (en) * | 2022-01-19 | 2023-08-01 | 腾讯科技(深圳)有限公司 | Information prediction method, device, electronic equipment and storage medium |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180068371A1 (en) * | 2016-09-08 | 2018-03-08 | Adobe Systems Incorporated | Learning Vector-Space Representations of Items for Recommendations using Word Embedding Models |
CN109543112A (en) * | 2018-11-30 | 2019-03-29 | 苏州大学 | A kind of sequence of recommendation method and device based on cyclic convolution neural network |
-
2019
- 2019-05-14 CN CN201910398877.3A patent/CN110119467B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180068371A1 (en) * | 2016-09-08 | 2018-03-08 | Adobe Systems Incorporated | Learning Vector-Space Representations of Items for Recommendations using Word Embedding Models |
CN109543112A (en) * | 2018-11-30 | 2019-03-29 | 苏州大学 | A kind of sequence of recommendation method and device based on cyclic convolution neural network |
Cited By (66)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110633789A (en) * | 2019-08-27 | 2019-12-31 | 苏州市职业大学 | Self-attention network information processing method for streaming media recommendation |
WO2021038592A2 (en) | 2019-08-30 | 2021-03-04 | Tata Consultancy Services Limited | System and method for handling popularity bias in item recommendations |
EP4022553A4 (en) * | 2019-08-30 | 2023-08-23 | Tata Consultancy Services Limited | System and method for handling popularity bias in item recommendations |
CN110795618A (en) * | 2019-09-12 | 2020-02-14 | 腾讯科技(深圳)有限公司 | Content recommendation method, device, equipment and computer readable storage medium |
CN110795618B (en) * | 2019-09-12 | 2024-05-14 | 腾讯科技(深圳)有限公司 | Content recommendation method, device, equipment and computer readable storage medium |
CN110619082B (en) * | 2019-09-20 | 2022-05-17 | 苏州市职业大学 | Project recommendation method based on repeated search mechanism |
CN110717098A (en) * | 2019-09-20 | 2020-01-21 | 中国科学院自动化研究所 | Meta-path-based context-aware user modeling method and sequence recommendation method |
CN110619082A (en) * | 2019-09-20 | 2019-12-27 | 苏州市职业大学 | Project recommendation method based on repeated search mechanism |
CN110717098B (en) * | 2019-09-20 | 2022-06-24 | 中国科学院自动化研究所 | Meta-path-based context-aware user modeling method and sequence recommendation method |
CN110598118A (en) * | 2019-09-23 | 2019-12-20 | 腾讯科技(深圳)有限公司 | Resource object recommendation method and device and computer readable medium |
CN110705709A (en) * | 2019-10-14 | 2020-01-17 | 支付宝(杭州)信息技术有限公司 | Method and device for training neural network model of graph |
CN110705709B (en) * | 2019-10-14 | 2021-03-23 | 支付宝(杭州)信息技术有限公司 | Method and device for training neural network model of graph |
CN110807156A (en) * | 2019-10-23 | 2020-02-18 | 山东师范大学 | Interest recommendation method and system based on user sequence click behaviors |
CN110796313A (en) * | 2019-11-01 | 2020-02-14 | 北京理工大学 | Session recommendation method based on weighted graph volume and item attraction model |
CN110796313B (en) * | 2019-11-01 | 2022-05-31 | 北京理工大学 | Session recommendation method based on weighted graph volume and item attraction model |
CN110826700A (en) * | 2019-11-13 | 2020-02-21 | 中国科学技术大学 | Method for realizing and classifying bilinear graph neural network model for modeling neighbor interaction |
CN110826700B (en) * | 2019-11-13 | 2021-04-23 | 中国科学技术大学 | Method for realizing and classifying bilinear graph neural network model for modeling neighbor interaction |
CN111046257A (en) * | 2019-12-09 | 2020-04-21 | 北京百度网讯科技有限公司 | Session recommendation method and device and electronic equipment |
CN111046257B (en) * | 2019-12-09 | 2023-07-04 | 北京百度网讯科技有限公司 | Session recommendation method and device and electronic equipment |
JP7074964B2 (en) | 2019-12-09 | 2022-05-25 | ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド | Recommended session methods, equipment and devices |
JP2022516593A (en) * | 2019-12-09 | 2022-03-01 | 北京百度網訊科技有限公司 | Recommended session methods, equipment and devices |
WO2021114590A1 (en) * | 2019-12-09 | 2021-06-17 | 北京百度网讯科技有限公司 | Session recommendation method and apparatus, and electronic device |
CN111179031B (en) * | 2019-12-23 | 2023-09-26 | 第四范式(北京)技术有限公司 | Training method, device and system for commodity recommendation model |
CN111179031A (en) * | 2019-12-23 | 2020-05-19 | 第四范式(北京)技术有限公司 | Training method, device and system for commodity recommendation model |
CN111259256B (en) * | 2020-02-13 | 2023-09-12 | 深圳市雅阅科技有限公司 | Content processing method, content processing device, computer readable storage medium and computer equipment |
CN111259256A (en) * | 2020-02-13 | 2020-06-09 | 腾讯科技(深圳)有限公司 | Content processing method, content processing device, computer-readable storage medium and computer equipment |
CN111242748B (en) * | 2020-02-21 | 2023-04-07 | 腾讯科技(深圳)有限公司 | Method, apparatus, and storage medium for recommending items to a user |
CN111242748A (en) * | 2020-02-21 | 2020-06-05 | 腾讯科技(深圳)有限公司 | Method, apparatus, and storage medium for recommending items to a user |
CN111813640A (en) * | 2020-03-25 | 2020-10-23 | 北京嘀嘀无限科技发展有限公司 | Method, apparatus, electronic device and storage medium for predicting user behavior |
WO2021196954A1 (en) * | 2020-03-30 | 2021-10-07 | 北京沃东天骏信息技术有限公司 | Serialized data processing method and device, and text processing method and device |
CN111461301A (en) * | 2020-03-30 | 2020-07-28 | 北京沃东天骏信息技术有限公司 | Serialized data processing method and device, and text processing method and device |
CN111460331B (en) * | 2020-04-07 | 2020-12-25 | 中国计量大学 | Conversation recommendation method based on space-time diagram |
CN111460331A (en) * | 2020-04-07 | 2020-07-28 | 中国计量大学 | Conversation recommendation method based on space-time diagram |
CN111667067A (en) * | 2020-05-28 | 2020-09-15 | 平安医疗健康管理股份有限公司 | Recommendation method and device based on graph neural network and computer equipment |
CN111538906A (en) * | 2020-05-29 | 2020-08-14 | 支付宝(杭州)信息技术有限公司 | Information pushing method and device based on privacy protection |
CN111538906B (en) * | 2020-05-29 | 2023-06-20 | 支付宝(杭州)信息技术有限公司 | Information pushing method and device based on privacy protection |
CN111708876A (en) * | 2020-06-16 | 2020-09-25 | 北京百度网讯科技有限公司 | Method and device for generating information |
CN111708876B (en) * | 2020-06-16 | 2023-07-18 | 北京百度网讯科技有限公司 | Method and device for generating information |
CN112150210A (en) * | 2020-06-19 | 2020-12-29 | 南京理工大学 | Improved neural network recommendation method and system based on GGNN (global warming network) |
CN111737578B (en) * | 2020-06-22 | 2024-04-02 | 陕西师范大学 | Recommendation method and system |
CN111737578A (en) * | 2020-06-22 | 2020-10-02 | 陕西师范大学 | Recommendation method and system |
CN111859160A (en) * | 2020-08-07 | 2020-10-30 | 成都理工大学 | Method and system for recommending session sequence based on graph neural network |
CN111949865A (en) * | 2020-08-10 | 2020-11-17 | 杭州电子科技大学 | Interest point recommendation method based on graph neural network and user long-term and short-term preference |
CN112084404B (en) * | 2020-09-01 | 2024-03-01 | 北京百度网讯科技有限公司 | Content recommendation method, device, equipment and medium |
CN112084404A (en) * | 2020-09-01 | 2020-12-15 | 北京百度网讯科技有限公司 | Content recommendation method, device, equipment and medium |
CN112258262B (en) * | 2020-09-15 | 2023-09-26 | 浙江大学 | Session recommendation method based on convolution self-attention network |
CN112258262A (en) * | 2020-09-15 | 2021-01-22 | 浙江大学 | Conversation recommendation method based on convolution self-attention network |
CN112256879A (en) * | 2020-10-29 | 2021-01-22 | 贝壳技术有限公司 | Information processing method and apparatus, electronic device, and computer-readable storage medium |
CN112541639A (en) * | 2020-12-22 | 2021-03-23 | 宜宾电子科技大学研究院 | Recommendation system scoring prediction method based on graph neural network and attention mechanism |
CN112541639B (en) * | 2020-12-22 | 2023-07-07 | 宜宾电子科技大学研究院 | Recommendation system scoring prediction method based on graph neural network and attention mechanism |
CN112559878A (en) * | 2020-12-24 | 2021-03-26 | 山西大学 | Sequence recommendation system and recommendation method based on graph neural network |
CN112579914A (en) * | 2020-12-31 | 2021-03-30 | 安徽大学 | Time perception service recommendation system and method based on self-attention factorization machine |
CN112733018B (en) * | 2020-12-31 | 2022-12-06 | 哈尔滨工程大学 | Session recommendation method based on graph neural network GNN and multi-task learning |
CN112733018A (en) * | 2020-12-31 | 2021-04-30 | 哈尔滨工程大学 | Session recommendation method based on graph neural network GNN and multi-task learning |
CN112685657B (en) * | 2021-01-06 | 2022-03-04 | 中国计量大学 | Conversation social recommendation method based on multi-mode cross fusion graph network |
CN112685657A (en) * | 2021-01-06 | 2021-04-20 | 中国计量大学 | Conversation social recommendation method based on multi-mode cross fusion graph network |
CN112765461A (en) * | 2021-01-12 | 2021-05-07 | 中国计量大学 | Session recommendation method based on multi-interest capsule network |
CN112966193B (en) * | 2021-03-05 | 2023-07-25 | 北京百度网讯科技有限公司 | Travel intention deducing method, model training method, related device and electronic equipment |
CN112966193A (en) * | 2021-03-05 | 2021-06-15 | 北京百度网讯科技有限公司 | Travel intention inference method, model training method, related device and electronic equipment |
CN113467740A (en) * | 2021-07-15 | 2021-10-01 | 青岛博天数通信息科技有限公司 | Video monitoring array display optimization method and device based on joint coding |
CN113467740B (en) * | 2021-07-15 | 2024-02-02 | 青岛博天数通信息科技有限公司 | Video monitoring array display optimization method and device based on joint coding |
WO2023108324A1 (en) * | 2021-12-13 | 2023-06-22 | 中国科学院深圳先进技术研究院 | Comparative learning enhanced two-stream model recommendation system and algorithm |
CN116521972A (en) * | 2022-01-19 | 2023-08-01 | 腾讯科技(深圳)有限公司 | Information prediction method, device, electronic equipment and storage medium |
CN116521972B (en) * | 2022-01-19 | 2024-07-12 | 腾讯科技(深圳)有限公司 | Information prediction method, device, electronic equipment and storage medium |
CN114743053A (en) * | 2022-04-14 | 2022-07-12 | 电子科技大学 | Magnetic resonance image auxiliary processing system based on graph neural network and self attention |
CN114743053B (en) * | 2022-04-14 | 2023-04-25 | 电子科技大学 | Magnetic resonance image auxiliary processing system based on graph neural network and self-attention |
Also Published As
Publication number | Publication date |
---|---|
CN110119467B (en) | 2023-05-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110119467A (en) | A kind of dialogue-based item recommendation method, device, equipment and storage medium | |
Sun et al. | What and how: generalized lifelong spectral clustering via dual memory | |
WO2021078027A1 (en) | Method and apparatus for constructing network structure optimizer, and computer-readable storage medium | |
Singer et al. | Detecting memory and structure in human navigation patterns using Markov chain models of varying order | |
EP3992857A1 (en) | Method and device for generating neural network model, and computer-readable storage medium | |
CN111797321B (en) | Personalized knowledge recommendation method and system for different scenes | |
CN110083770B (en) | Sequence recommendation method based on deeper feature level self-attention network | |
CN111541570B (en) | Cloud service QoS prediction method based on multi-source feature learning | |
CN110097433B (en) | Attention mechanism-based recommendation method, device, equipment and storage medium | |
CN112115352B (en) | Session recommendation method and system based on user interests | |
CN110516160A (en) | User modeling method, the sequence of recommendation method of knowledge based map | |
CN111428147A (en) | Social recommendation method of heterogeneous graph volume network combining social and interest information | |
CN110168578A (en) | Multitask neural network with task particular path | |
CN110781409B (en) | Article recommendation method based on collaborative filtering | |
CN106056209B (en) | A kind of inquiry lexical item weight learning method based on Recognition with Recurrent Neural Network | |
CN110263245B (en) | Method and device for pushing object to user based on reinforcement learning model | |
CN111737578A (en) | Recommendation method and system | |
CN110795618B (en) | Content recommendation method, device, equipment and computer readable storage medium | |
CN115270007B (en) | POI recommendation method and system based on mixed graph neural network | |
CN111695024A (en) | Object evaluation value prediction method and system, and recommendation method and system | |
CN113590976A (en) | Recommendation method of space self-adaptive graph convolution network | |
CN114925270B (en) | Session recommendation method and model | |
CN109753275A (en) | Recommended method, device, storage medium and the electronic equipment of Application Programming Interface | |
CN116992151A (en) | Online course recommendation method based on double-tower graph convolution neural network | |
CN110837567A (en) | Method and system for embedding knowledge graph |
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 |