CN110060097A - User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks - Google Patents
User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks Download PDFInfo
- Publication number
- CN110060097A CN110060097A CN201910257110.9A CN201910257110A CN110060097A CN 110060097 A CN110060097 A CN 110060097A CN 201910257110 A CN201910257110 A CN 201910257110A CN 110060097 A CN110060097 A CN 110060097A
- Authority
- CN
- China
- Prior art keywords
- user
- term preference
- project
- vector
- long
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 57
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 29
- 230000007246 mechanism Effects 0.000 title claims abstract description 29
- 230000003993 interaction Effects 0.000 claims abstract description 64
- 230000007774 longterm Effects 0.000 claims abstract description 52
- 230000003542 behavioural effect Effects 0.000 claims abstract description 31
- 238000004364 calculation method Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 25
- 239000011159 matrix material Substances 0.000 claims description 16
- 238000003780 insertion Methods 0.000 claims description 14
- 230000037431 insertion Effects 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 7
- 230000002452 interceptive effect Effects 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 238000012512 characterization method Methods 0.000 claims description 4
- 239000000203 mixture Substances 0.000 claims description 4
- 238000011176 pooling Methods 0.000 claims description 4
- 241001269238 Data Species 0.000 claims description 3
- 238000012549 training Methods 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 230000006399 behavior Effects 0.000 description 15
- 238000013528 artificial neural network Methods 0.000 description 7
- 238000013507 mapping Methods 0.000 description 4
- 238000011161 development Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 235000014121 butter Nutrition 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 235000013336 milk Nutrition 0.000 description 2
- 239000008267 milk Substances 0.000 description 2
- 210000004080 milk Anatomy 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 241000287196 Asthenes Species 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000005284 excitation Effects 0.000 description 1
- 235000013312 flour Nutrition 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06Q30/0241—Advertisements
- G06Q30/0251—Targeted advertisements
- G06Q30/0255—Targeted advertisements based on user history
-
- 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
- G06Q30/0241—Advertisements
- G06Q30/0277—Online advertisement
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- General Business, Economics & Management (AREA)
- Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Marketing (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks that the invention discloses a kind of, wherein the following steps are included: acquiring the passing behavioral data of user as long-term preference interaction sequence data；The weight of each behavioral data in the long-term preference interaction sequence data is calculated using attention mechanism algorithm, and goes out the long-term preference of user according to the weight calculation；The behavioral data in a period of time is chosen as short-term preference interaction sequence data；According to the short-term preference interaction sequence data, the short-term preference of user is calculated using convolutional neural networks algorithm；The long-term preference of user is spliced with short-term preference, to generate user behavior sequence of recommendation；The present invention can recommend compared to the method for the interaction between traditional fitting user-project or project-project, the long-term preference and short-term preference of enough combination users quickly and effectively to user.
Description
Technical field
The present invention relates to user behavior sequence of recommendation methods, are based on attention mechanism and volume more specifically to one kind
The user behavior sequence of recommendation method of product neural network.
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.Happiness of the recommender system by prediction user to information resources
Good degree carries out information filtering.
With the fast development of platform economy, Amazon, many companies such as Taobao and excellent step are creating self ecosystem
System, by keeping user here with interacting for products & services.User can access these by mobile device in daily life
Platform can thus generate a large amount of user behaviors log.For example, in June, 2017, excellent step has accumulated 6800 general-purpose families travelling number
According to；Registering more than 11,000,000,000 has been produced more than 5000 general-purpose families on Foursquare.Effective recommender system is constructed,
Key factor is accurately to characterize and understand the interest and taste of user, and these are changed and development continuous always.In order to
Realize the target, sequence of recommendation comes into being, with the item that may be interacted based on user's history active sequences come recommended user
Mesh.
It is different from traditional recommender system, in sequence of recommendation, transaction or session of more and more data from series.
These transaction or session form the sequence pattern of user, and the project of the next possible access of user is more likely to depend on user most
The nearly project for generating interaction.User is caused next to buy the probability of flour higher than independent for example, buying milk and butter together
Buy the probability of milk or butter.However, traditional recommended method, such as collaborative filtering, matrix decomposition and Top-N recommend, it is uncomfortable
Capture sequence pattern is closed, because they will not model the sequence of movement.In order to solve this problem, early stage based on horse
The sequence method of Markov's chain generallys use individual model to characterize the long-term preference of user and short-term preference, then by them
It integrates.However, the partial order behavior between these Method Modeling every two adjacencies based on Markov chain, but
It is difficult to model higher order relationship.In recent years, deep neural network has obtained in-depth study in related fields, generates to sequence of recommendation
Very big influence.Most popular neural network for modeling user's sequence pattern is Recognition with Recurrent Neural Network (RNN).It is based on
The method of RNN is more powerful than traditional sequential grammar.However, RNN assumes that time dependence is monotonically changed, it means that it is current or
Hidden state is more important than previous.In sequence of recommendation, and not all adjacent movement all has dependence.In order to solve this
A problem has scholar to propose and carries out convolution to the insertion vector of project using the convolution filter of convolutional neural networks (CNN)
Operate and then obtain the sequence signature of part.However, this method is only using user's embeded matrix as the long-term preference of user,
This is not enough to learn the long-term preference of user.
Therefore, we propose a kind of sequence of recommendation side based on attention mechanism and convolutional neural networks in the present invention
Method.Attention mechanism is used to model the long-term preference of user, and convolutional neural networks are used to excavate the short-term sequence preference of user, most
The two is recommended in conjunction with to generate afterwards.
Summary of the invention
It is excellent it is an object of the invention to solve at least the above problems and/or defect, and provide at least to will be described later
Point.
The present invention provides the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks, energy
The long-term preference and short-term preference of enough combination users quickly and effectively is recommended to user.
In order to realize these purposes according to the present invention and other advantages, provide a kind of based on attention mechanism and convolution
The user behavior sequence of recommendation method of neural network, wherein the following steps are included: acquiring the passing behavioral data of user as length
Phase preference interaction sequence data；
The weight of each behavioral data in the long-term preference interaction sequence data, and root are calculated using attention mechanism algorithm
Go out the long-term preference of user according to the weight calculation；
The behavioral data in a period of time is chosen as short-term preference interaction sequence data；
According to the short-term preference interaction sequence data, the short-term inclined of user is calculated using convolutional neural networks algorithm
It is good；
The long-term preference of user is spliced with short-term preference, to generate user behavior sequence of recommendation.
Further, the passing behavioral data is all passing behavioral datas of the user on a fixed platform, or
For the passing behavioral data for calculating certain time to the past from current date.
Further, the behavioral data in described selection a period of time is to be selected from current date to rise to the past calculating centainly
The passing behavioral data of time.
Further, the method for the short-term preference that user is calculated using convolutional neural networks algorithm includes following step
It is rapid:
Using the short-term preference interaction sequence data of horizontal convolution operation processing, vector o is obtained；
Short-term preference interaction sequence data are handled using vertical convolution operation, obtain vector
Link vector o and vectorThe short-term preference of user is calculated with this.
Further, the splicing is input in a fully-connected network, to generate user behavior sequence of recommendation.
Further, described to calculate each behavior number in the long-term preference interaction sequence data using attention mechanism algorithm
According to weight the following steps are included:
Calculate the weight α of each behavioral data in the long-term preference interaction sequence dataj, comprising:
It calculates to hide first and indicates hj
Wherein,It is the activation primitive of line rectification function (RELU), W1And b1For empirical parameter, vjLong-term preference is indicated to hand over
J-th object vector is represented in mutual sequence data；
Then h is indicated according to hidingjCalculate weight αj
Wherein, u is the vector of the current user to be calculated of representative, uTIndicate the transposition of the vector, LuRepresent the long-term of the user
Preference interaction sequence data.
Further, it is described according to the weight calculation go out user long-term preference the following steps are included:
The project for considering weight is embedded in the long-term preference of the summation characterization user of long-term preference interaction sequence data
ulong:
Further, the short-term preference interaction sequence data be in t moment, by the insertion vector of the l project of the past,
I.e. from vt-lTo vt-1, the matrix E of composition(u, t)It is considered as image, carries out horizontal and vertical convolution algorithm respectively；
It is described utilize the short-term preference interaction sequence data of horizontal convolution operation processing, obtain vector o the following steps are included:
Utilize horizontal filterIt is slipped over simultaneously from project matrix in different moments
Sequence signature is captured, n indicates number of filter, and h indicates the height of horizontal filter, and k indicates the dimension of vector, for each
Moment t, with each FxFrom E(u, t)Top slide to bottom, a value of the i-th ∈ [1, l-h+1] of the result of feature interaction by with
Lower formula calculates,
Wherein,It is the activation primitive of line rectification function (RULU), indicates inner product operation,Representing matrix E(u, t)The the i-th to the i-th+h-1 row, so, final convolution results are expressed as vector cx
Then, to the output result c of single filterxMaximum pond (max-pooling) is carried out to operate to obtain maximum
Value, therefore, for the last output result vector of n filterIt indicates,
O=[max (c1), max (c2) ... max (en)]
Wherein,
It is described to handle short-term preference interaction sequence data using vertical convolution operation, obtain vectorThe following steps are included:
WithA filterFilterWith E(u, t)Each column interact, convolution
As a result it is I.e.
P is E(u,t)Pth row polymerize the insertion vector of front project using vertical filter,The output of a filter
As a result it is
Further, the link vector o and vectorMethod are as follows:
Connection result
For parameter,For parameter,It is activation primitive.
Further, the method that the long-term preference of user is spliced with short-term preference includes:
Splicing result y is obtained by following formula(u,t):
For parameter,For parameter, N indicates the quantity of project, numerical valueTable
Show that, in the probability possibility size that moment t user u is interacted with project s, v is the passing project of the user；
Wherein parameter W2, b1, b2, W1, W3, b3For empirical value, or obtained according to following methods:
Using gradient ascent algorithm and back-propagation method come Optimal Parameters:
It is assumed that user is more partial to the project of interaction of lower a period of time relative to unobservable project, then define
Such a ordering ruleFor project p and q, then have:P is user u in t
The project of interaction is carved, q is that user u in t moment does not generate interactive project also, for { (u, the L observed every timeu, p) }, it produces
Raw a pair of Preference order D={ (u, Lu, p, q) },
Then, by maximizing following objective function come training pattern,
For Θ representative model to all parameters, λ is regularization weight, and δ is logistic function,
By optimization process above, the optimal solution of model parameter Θ is acquired, all parameters are released according to optimal solution.
The method that the present invention compares the interaction between traditional fitting user-project or project-project, it is enough fast and effective
Combination user long-term preference and short-term preference to user recommend.
Further advantage, target and feature of the invention will be partially reflected by the following instructions, and part will also be by this
The research and practice of invention and be understood by the person skilled in the art.
Detailed description of the invention
Fig. 1 is the mould of the user behavior sequence of recommendation method of the present invention based on attention mechanism and convolutional neural networks
Type figure.
Specific embodiment
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.
It should be appreciated that such as " having ", "comprising" and " comprising " term used herein do not allot one or more
The presence or addition of a other elements or combinations thereof.
As shown in Figure 1, the user behavior sequence of recommendation side of the present invention based on attention mechanism and convolutional neural networks
Method, wherein the following steps are included: acquiring the passing behavioral data of user as long-term preference interaction sequence data；
The weight of each behavioral data in the long-term preference interaction sequence data, and root are calculated using attention mechanism algorithm
Go out the long-term preference of user according to the weight calculation；
The behavioral data in a period of time is chosen as short-term preference interaction sequence data；
According to the short-term preference interaction sequence data, the short-term inclined of user is calculated using convolutional neural networks algorithm
It is good；
The long-term preference of user is spliced with short-term preference, to generate user behavior sequence of recommendation.
Further, the passing behavioral data is all passing behavioral datas of the user on a fixed platform, or
For the passing behavioral data for calculating certain time to the past from current date.
Further, the behavioral data in described selection a period of time is to be selected from current date to rise to the past calculating centainly
The passing behavioral data of time.
Further, the method for the short-term preference that user is calculated using convolutional neural networks algorithm includes following step
It is rapid:
Using the short-term preference interaction sequence data of horizontal convolution operation processing, vector o is obtained；
Short-term preference interaction sequence data are handled using vertical convolution operation, obtain vector
Link vector o and vectorThe short-term preference of user is calculated with this.
Further, the splicing is input in a fully-connected network, to generate user behavior sequence of recommendation.
Further, described to calculate each behavior number in the long-term preference interaction sequence data using attention mechanism algorithm
According to weight the following steps are included:
Calculate the weight α of each behavioral data in the long-term preference interaction sequence dataj, comprising:
It calculates to hide first and indicates hj
Wherein,It is the activation primitive of line rectification function (RELU), W1And b1For empirical parameter, vjLong-term preference is indicated to hand over
J-th object vector is represented in mutual sequence data；
Then h is indicated according to hidingjCalculate weight αj
Wherein, u is the vector of the current user to be calculated of representative, uTIndicate the transposition of the vector, LuRepresent the long-term of the user
Preference interaction sequence data.
Further, it is described according to the weight calculation go out user long-term preference the following steps are included:
The project for considering weight is embedded in the long-term preference of the summation characterization user of long-term preference interaction sequence data
ulong:
。
Further, the short-term preference interaction sequence data be in t moment, by the insertion vector of the l project of the past,
I.e. from vt-lTo vt-1, the matrix E of composition(u,t)It is considered as image, carries out horizontal and vertical convolution algorithm respectively；
It is described utilize the short-term preference interaction sequence data of horizontal convolution operation processing, obtain vector o the following steps are included:
Utilize horizontal filterIt is slipped over simultaneously from project matrix in different moments
Sequence signature is captured, n indicates number of filter, and h indicates the height of horizontal filter, and k indicates the dimension of vector, for each
Moment t, with each FxFrom E(u,t)Top slide to bottom, a value of the i-th ∈ [1, l-h+1] of the result of feature interaction by with
Lower formula calculates,
Wherein,It is the activation primitive of line rectification function (RULU), indicates inner product operation,Representing matrix
E(u,t)The the i-th to the i-th+h-1 row, so, final convolution results are expressed as vector cx
Then, to the output result c of single filterxMaximum pond (max-pooling) is carried out to operate to obtain maximum
Value, therefore, for the last output result vector of n filterIt indicates,
O=[max (c1),max(c2),...max(cn)]
Wherein,
It is described to handle short-term preference interaction sequence data using vertical convolution operation, obtain vectorThe following steps are included: withA filterFilterWithEach column interact, convolution results are I.e.
P is E(u,t)Pth row polymerize the insertion vector of front project using vertical filter,The output of a filter
As a result it is
Further, the connection link vector o and vectorMethod are as follows:
Connection result
For parameter,For parameter,It is activation primitive.
Further, the method that the long-term preference of user is spliced with short-term preference includes:
Splicing result y is obtained by following formula(u,t):
For parameter,For parameter, N indicates the quantity of project, numerical valueTable
Show that, in the probability possibility size that moment t user u is interacted with project s, v is the passing project of the user；
Wherein parameter W2, b1, b2, W1, W3, b3For empirical value, or obtained according to following methods:
Using gradient ascent algorithm and back-propagation method come Optimal Parameters:
It is assumed that user is more partial to the project of interaction of lower a period of time relative to unobservable project, then define
Such a ordering ruleFor project p and q, then have:P is user u in t
The project of interaction is carved, q is that user u in t moment does not generate interactive project also, for { (u, the L observed every timeu, p) }, it produces
Raw a pair of Preference order D={ (u, Lu, p, q) },
Then, by maximizing following objective function come training pattern,
For Θ representative model to all parameters, λ is regularization weight, and δ is logistic function,
By optimization process above, the optimal solution of model parameter Θ is acquired, all parameters are released according to optimal solution.
Embodiment
Firstly, we use U={ u1,u2,……,uMIndicate M user set, V={ v1, v2... ..., vNIndicate source
The N number of project set in domain.Each user generates interaction sequence with a series of projectsThis sequence arranges sequentially in time, and wherein t indicates opposite time index
Rather than the absolute time, | Lu| indicate the interaction sequence length of user.There is the above symbol to indicate, our sequence of recommendation is appointed
Business is defined as follows, we are absorbed in from implicit feedback and user's interaction sequence feedback data (for example, user's continuously registers and purchase
Buy transaction record) in extract information.As the history interaction sequence L of given a user u and heu, our purpose is from interaction
In record excavate user long-term preference and short-term preference come recommended user u next a period of time in there may be interactive items
Mesh set.
Correlation technique
(1) gradient ascent algorithm
Gradient rise method is more commonly used optimization algorithm.Gradient rise method is based on the idea that be: finding certain function most
Big value, the best way is sought along the gradient direction of the function.If gradient is denoted asThen the gradient of function f (x, y) can
It indicates are as follows:
This gradient means to move along the direction of xIt moves in direction along yWherein, function f (x, y)
It must be defined on point to be calculated and can be micro-.In each iteration, gradient ascent algorithm all can move one along gradient direction
Step.Wherein, gradient operator always points at the fastest-rising direction of functional value.Mentioned here is moving direction, without referring to
Amount of movement size.The magnitude is known as step-length, is denoted as α.The iterative formula that gradient rises can indicate are as follows:
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) back-propagation algorithm
Back-propagation algorithm, i.e. BP algorithm are to be suitable for a kind of study of multilayer neural networks in the case where there is tutor's guidance
Algorithm, it is established on the basis of gradient descent method.The input/output relation of BP network is substantially a kind of mapping relations: one
The function completed of BP neural network of input m output be tieed up from n Euclidean space into m dimension Euclidean space one finite field it is continuous
Mapping, this mapping have nonlinearity.Its information processing capability is multiple compound from simple non-linear functions, because
This reappears ability with very strong function.This is the basis that BP algorithm is able to apply.Back-propagation algorithm is mainly by two links
(excitation is propagated, weight updates) iterative cycles iteration, until the response to input of network reaches scheduled target zone.
The learning process of BP algorithm is made of forward-propagating process and back-propagation process.It is defeated during forward-propagating
Enter information by input layer through hidden layer, successively handles and be transmitted to output layer.If cannot get desired output valve in output layer,
The quadratic sum of output and desired error is then taken to be transferred to backpropagation as objective function, successively find out objective function to each mind
Partial derivative through first weight constitutes objective function and measures to the ladder of weight vector, and as the foundation of modification weight, the study of network exists
It is completed during weight modification.When error reaches desired value, e-learning terminates.
As shown in Figure 1, proposing a kind of completely new sequence of recommendation method based on attention mechanism and convolutional neural networks.
It is proposed that model be broadly divided into 4 layers, first layer is embeding layer (Embedding Layer), and user and item are mapped to
Continuous lower dimensional space；The second layer is the long-term preference that attention layer (Attention Layer) is used to model user；Third layer
It is convolutional layer (Convolutional Layer), for excavating the short-term sequence preference of user；4th layer is full articulamentum
(Fully-connected Layer) is fitted the long-term preference of user and short-term preference and recommends to generate.The experimental results showed that I
The method that proposes better than state-of-the-art method.
1、Embedding Layer
It is indicated similar to the discrete vocabulary symbol in natural language processing, the expression ability of the ID of original user and item
It is limited.Therefore, user and item are mapped in two low-dimensional continuous spaces by our model first step, are denoted as respectivelyMatrix UeAnd matrix VeRespectively indicate user_embedding and item_ that mapping obtains
Embedding, K indicate the dimension of lower dimensional space, M and N not Biao Shi user and project quantity.
2、Attention Layer
Since the long-term project collection of user usually changes over time, the long-term preference of static state for learning each user is indicated
The dynamic of the long-term preference of user cannot be expressed completely.Furthermore, it is believed that identical project may produce different users
Raw different influence, and different projects can generate different influences to that will access or buy down several projects.Example
Such as, user a buys article m because of the interest of oneself, and user b purchase article m is to give friend.In such case
Under, predicting what when they may buy in the near future, project m there should be different weights to user a and b.
In order to meet above-mentioned requirements, we use attention mechanism, which has been successfully applied to many tasks, such as answer automatically
It answers, document classification and recommendation.Given user u, we calculate his/her historical series L firstuIn each project importance,
Then the insertion vector that we polymerize these projects is indicated with the long-term preference for forming user.In form, attention net definitions
It is as follows:
It is parameter, vjThe insertion vector of expression project j,It is RELU activation primitive, is used to
The non-thread sexuality of lift scheme.Firstly, we are by the history interaction sequence L of user uuEach of project lower dimensional space
Vector vjIt is input in a fully-connected network and obtains its hiding expression hj；Then, its corresponding attention is acquired according to formula (1)
Power weight αj.It is that each input uses identical context vector in traditional attention model, we regard the insertion vector of user u
For context vector, and it will pay attention to score αjAs hjNormalization similarity between u measures together with softmax function,
Project j is characterized to the importance u of user.Finally, the summation characterization for the project insertion vector that we are weighted with attention score
The long-term preference of user, is denoted as ulong:
3, convolutional layer
Convolutional layer utilizes the local sequence signature for going capture user of the convolution filter (filter) of CNN, for given
A user u and its interaction sequence Lu, we are in t moment, by the insertion vector of l project of front (i.e. from vt-lTo vt-1)
Composition matrix E (u,t) it is considered as " image ", horizontal and vertical convolution algorithm is carried out respectively.Horizontal convolutional layer
In horizontal convolution operation, we utilize horizontal filterIn difference
Moment slips over from project matrix and captures sequence signature, and n indicates number of filter, and h indicates the height of horizontal filter.For
Each moment t, we are with each FxFrom E(u,t)Top slide to bottom.I-th ∈ [1, l-h+ of the result of feature interaction
1] a value is calculated by formula (2),
It is RULU activation primitive, indicates inner product operation,Representing matrix E(u,t)The the i-th to the i-th+h-1 row.
So final convolution results are expressed as vector cx
Then, we are to the output result c to single filterxMaximum pond (max-pooling) is carried out to operate to obtain
Obtain maximum value.Therefore, for the last output result of n filter, we use vectorIt indicates,
O=[max (c1),max(c2) ... max (cn)]
Vertical convolutional layer
In vertical convolutional layer, we are with (~) come label symbol.
It is similar with horizontal convolutional layer, Wo MenyongA filterFilterWith E(u , t)Each column interact, convolution results are
I.e.
P is E(u, t)Pth row.Using vertical filter, we can polymerize the insertion vector of front project, therefore, this
The local sequence that a little vertical filters can capture user u with the weighted sum of the insertion vector of l project before t moment is special
Sign.The output result of a filter is
It is not necessary to be operated to vertical convolution results using pondization, because it is desirable that keeping the poly- of each potential dimension
It closes.Therefore, the output of this layer is
Finally, we connect the output of two convolutional layers, then they are input in a fully-connected network, to obtain
More advanced and abstract sequence signature indicates
It is activation primitive.
4、Fully-connected Layer
In order to obtain the overall preference of user, we will also be noted that the output of power layer and the output of convolutional layer are connected to one
It rises, and they is input in a full connection output layer, be expressed as follows,
N indicates the quantity of project, numerical valueIt indicates to use in moment t
The probability possibility size that family u is interacted with project s.Finally, according to y(u, t)Carry out the project that recommended user next may be interactive.
Network model parameter optimization
For Optimized model, optimal model parameter is obtained, we are using based on above-mentioned gradient ascent algorithm and reversely
Transmission method carrys out Optimized model.Meanwhile we use BRP Optimality Criteria objective function.We assume that relative to unobservable
Project, user are more partial to the project of purchase of lower a period of time, and then we define such a ordering ruleIt is right
In project p and q, then have:P is project of the user u in t moment interaction, and q is that user u exists
T moment does not generate interactive project also.For { (u, the L observed every timeu, p) }, we generate a pair of of Preference order D=
{(u,Lu,p,q)}。
Then, we train our model by the following objective function of maximization,
For Θ representative model to all parameters, λ is regularization weight, and δ is logistic function.
By optimization process above, we can be in the hope of the optimal solution of model parameter Θ.
Although the embodiments of the present invention have been disclosed as above, but it is not limited in listed fortune in specification and embodiments
With.It can be applied to various suitable the field of the invention completely.It for those skilled in the art, can be easily real
Now other modification.Therefore without departing from the general concept defined in the claims and the equivalent scope, the present invention is not limited to
Specific details and legend shown and described herein.
Claims (10)
1. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks, which is characterized in that including following
Step: the passing behavioral data of acquisition user is as long-term preference interaction sequence data；
The weight of each behavioral data in the long-term preference interaction sequence data is calculated using attention mechanism algorithm, and according to institute
State the long-term preference that weight calculation goes out user；
The behavioral data in a period of time is chosen as short-term preference interaction sequence data；
According to the short-term preference interaction sequence data, the short-term preference of user is calculated using convolutional neural networks algorithm；
The long-term preference of user is spliced with short-term preference, to generate user behavior sequence of recommendation.
2. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as described in claim 1,
It is characterized in that, the passing behavioral data is all passing behavioral datas of the user on a fixed platform, or for from currently
The passing behavioral data of certain time is calculated from date to the past.
3. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as described in claim 1,
It is characterized in that, the behavioral data chosen in a period of time is to be selected from current date to play the mistake for calculating certain time to the past
Toward behavioral data.
4. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as described in claim 1,
Be characterized in that, the method for the short-term preference that user is calculated using convolutional neural networks algorithm the following steps are included:
Using the short-term preference interaction sequence data of horizontal convolution operation processing, vector o is obtained；
Short-term preference interaction sequence data are handled using vertical convolution operation, obtain vector
Link vector o and vectorThe short-term preference of user is calculated with this.
5. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as described in claim 1,
It is characterized in that, the splicing is input in a fully-connected network, to generate user behavior sequence of recommendation.
6. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as described in claim 1,
It is characterized in that, the weight that each behavioral data in the long-term preference interaction sequence data is calculated using attention mechanism algorithm
The following steps are included:
Calculate the weight α of each behavioral data in the long-term preference interaction sequence dataj, comprising:
It calculates to hide first and indicates hj
Wherein,It is the activation primitive of line rectification function (RELU), W1And b1For empirical parameter, vjIndicate long-term preference interaction sequence
J-th object vector is represented in column data；
Then h is indicated according to hidingjCalculate weight αj
Wherein, u is the vector of the current user to be calculated of representative, uTIndicate the transposition of the vector, LuRepresent the long-term preference of the user
Interaction sequence data.
7. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as claimed in claim 6,
It is characterized in that, it is described according to the weight calculation go out user long-term preference the following steps are included:
The project for considering weight is embedded in the long-term preference u of the summation characterization user of long-term preference interaction sequence datalong:
8. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as claimed in claim 4,
It is characterized in that, the short-term preference interaction sequence data are in t moment, by the insertion vector of the l project of the past, i.e., from vt-l
To vt-1, the matrix E of composition(u,t)It is considered as image, carries out horizontal and vertical convolution algorithm respectively；
It is described utilize the short-term preference interaction sequence data of horizontal convolution operation processing, obtain vector o the following steps are included:
Utilize horizontal filterIt slips over and captures from project matrix in different moments
Sequence signature, n indicate number of filter, and h indicates the height of horizontal filter, and k indicates the dimension of vector, for each moment
T, with each FxFrom E(u,t)Top slide to bottom, a value of the i-th ∈ [1, l-h+1] of the result of feature interaction is by following public affairs
Formula calculates,
Wherein,It is the activation primitive of line rectification function (RULU), indicates inner product operation,Representing matrix E(u,t)
The the i-th to the i-th+h-1 row, so, final convolution results are expressed as vector cx
Then, to the output result c of single filterxMaximum pond (max-pooling) operation is carried out to obtain maximum value, because
This, for the last output result vector of n filterIt indicates,
O=[max (c1),max(c2),…max(cn)]
Wherein,
It is described to handle short-term preference interaction sequence data using vertical convolution operation, obtain vectorThe following steps are included:
WithA filterFilterWith E(u,t)Each column interact, convolution results are I.e.
P is E(u,t)Pth row polymerize the insertion vector of front project using vertical filter,The output result of a filter
For
9. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as claimed in claim 8,
It is characterized in that, the link vector o and vectorMethod are as follows:
Connection result
For parameter,For parameter,It is activation primitive.
10. the user behavior sequence of recommendation method based on attention mechanism and convolutional neural networks as described in claim 1,
It is characterized in that, the method that the long-term preference of user is spliced with short-term preference includes:
Splicing result y is obtained by following formula(u,t):
For parameter,For parameter, N indicates the quantity of project, numerical valueIt indicates
The probability possibility size that moment t user u is interacted with project s, v are the passing project of the user；
Wherein parameter W2, b1, b2, W1, W3, b3For empirical value, or obtained according to following methods:
Using gradient ascent algorithm and back-propagation method come Optimal Parameters:
It is assumed that user is more partial to the project of interaction of lower a period of time relative to unobservable project, then define in this way
One ordering ruleFor project p and q, then have:P is that user u is handed in t moment
Mutual project, q is that user u in t moment does not generate interactive project also, for { (u, the L observed every timeu, p) }, generate one
To Preference order D={ (u, Lu, p, q) },
Then, by maximizing following objective function come training pattern,
For Θ representative model to all parameters, λ is regularization weight, and δ is logistic function,
By optimization process above, the optimal solution of model parameter Θ is acquired, all parameters are released according to optimal solution.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910257110.9A CN110060097A (en) | 2019-04-01 | 2019-04-01 | User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910257110.9A CN110060097A (en) | 2019-04-01 | 2019-04-01 | User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110060097A true CN110060097A (en) | 2019-07-26 |
Family
ID=67318082
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910257110.9A Pending CN110060097A (en) | 2019-04-01 | 2019-04-01 | User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110060097A (en) |
Cited By (27)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245299A (en) * | 2019-06-19 | 2019-09-17 | 中国人民解放军国防科技大学 | Sequence recommendation method and system based on dynamic interaction attention mechanism |
| CN110377840A (en) * | 2019-07-29 | 2019-10-25 | 电子科技大学 | A kind of music list recommended method and system based on user's shot and long term preference |
| CN110555469A (en) * | 2019-08-15 | 2019-12-10 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
| CN110598847A (en) * | 2019-08-15 | 2019-12-20 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
| CN110619082A (en) * | 2019-09-20 | 2019-12-27 | 苏州市职业大学 | Project recommendation method based on repeated search mechanism |
| CN110619081A (en) * | 2019-09-20 | 2019-12-27 | 苏州市职业大学 | News pushing method based on interactive graph neural network |
| CN110659742A (en) * | 2019-09-10 | 2020-01-07 | 阿里巴巴集团控股有限公司 | Method and device for acquiring sequence representation vector of user behavior sequence |
| 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 |
| CN110851705A (en) * | 2019-10-09 | 2020-02-28 | 天津大学 | Project-based collaborative storage recommendation method and recommendation device thereof |
| CN110929164A (en) * | 2019-12-09 | 2020-03-27 | 北京交通大学 | Interest point recommendation method based on user dynamic preference and attention mechanism |
| CN111552881A (en) * | 2020-05-09 | 2020-08-18 | 苏州市职业大学 | Sequence recommendation method based on hierarchical variation attention |
| CN111582443A (en) * | 2020-04-22 | 2020-08-25 | 成都信息工程大学 | Recommendation method based on Mask mechanism and level attention mechanism |
| CN111753209A (en) * | 2020-07-02 | 2020-10-09 | 南京工业大学 | Sequence recommendation list generation method based on improved time sequence convolutional network |
| CN112016000A (en) * | 2020-07-20 | 2020-12-01 | 桂林电子科技大学 | Movie group recommendation method and system based on convolution collaborative filtering |
| CN112256971A (en) * | 2020-10-29 | 2021-01-22 | 清华大学深圳国际研究生院 | Sequence recommendation method and computer-readable storage medium |
| US10936950B1 (en) | 2019-08-15 | 2021-03-02 | Advanced New Technologies Co., Ltd. | Processing sequential interaction data |
| US10970350B2 (en) | 2019-08-15 | 2021-04-06 | Advanced New Technologies Co., Ltd. | Method and apparatus for processing user interaction sequence data |
| CN113076469A (en) * | 2020-01-03 | 2021-07-06 | 阿里巴巴集团控股有限公司 | Model construction method, data recommendation method and device |
| CN113448245A (en) * | 2021-04-14 | 2021-09-28 | 华南师范大学 | Deep learning-based dissolved oxygen control method and system in sewage treatment process |
| CN113468434A (en) * | 2021-09-06 | 2021-10-01 | 北京搜狐新动力信息技术有限公司 | Resource recommendation method, device, readable medium and equipment |
| CN113495991A (en) * | 2020-04-03 | 2021-10-12 | 北京沃东天骏信息技术有限公司 | Recommendation method and device |
| CN113868542A (en) * | 2021-11-25 | 2021-12-31 | 平安科技（深圳）有限公司 | Attention model-based push data acquisition method, device, equipment and medium |
| CN114154071A (en) * | 2021-12-09 | 2022-03-08 | 电子科技大学 | Emotion time sequence recommendation method based on attention mechanism |
| CN114169869A (en) * | 2022-02-14 | 2022-03-11 | 北京大学 | Attention mechanism-based post recommendation method and device |
| CN115062230A (en) * | 2022-08-16 | 2022-09-16 | 深圳市海清视讯科技有限公司 | Data processing method, apparatus, device, medium, and program product |
| CN115545834A (en) * | 2022-10-10 | 2022-12-30 | 杭州电子科技大学 | Personalized service recommendation method based on graph neural network and metadata |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108648049A (en) * | 2018-05-03 | 2018-10-12 | 中国科学技术大学 | A kind of sequence of recommendation method based on user behavior difference modeling |
| CN109299396A (en) * | 2018-11-28 | 2019-02-01 | 东北师范大学 | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model |
| CN109359140A (en) * | 2018-11-30 | 2019-02-19 | 苏州大学 | A kind of sequence of recommendation method and device based on adaptive attention |
| CN109544306A (en) * | 2018-11-30 | 2019-03-29 | 苏州大学 | A kind of cross-cutting recommended method and device based on user behavior sequence signature |
| 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
- 2019-04-01 CN CN201910257110.9A patent/CN110060097A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108648049A (en) * | 2018-05-03 | 2018-10-12 | 中国科学技术大学 | A kind of sequence of recommendation method based on user behavior difference modeling |
| CN109299396A (en) * | 2018-11-28 | 2019-02-01 | 东北师范大学 | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model |
| CN109359140A (en) * | 2018-11-30 | 2019-02-19 | 苏州大学 | A kind of sequence of recommendation method and device based on adaptive attention |
| CN109544306A (en) * | 2018-11-30 | 2019-03-29 | 苏州大学 | A kind of cross-cutting recommended method and device based on user behavior sequence signature |
| 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 (42)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110245299B (en) * | 2019-06-19 | 2022-02-08 | 中国人民解放军国防科技大学 | Sequence recommendation method and system based on dynamic interaction attention mechanism |
| CN110245299A (en) * | 2019-06-19 | 2019-09-17 | 中国人民解放军国防科技大学 | Sequence recommendation method and system based on dynamic interaction attention mechanism |
| CN110377840A (en) * | 2019-07-29 | 2019-10-25 | 电子科技大学 | A kind of music list recommended method and system based on user's shot and long term preference |
| CN110377840B (en) * | 2019-07-29 | 2020-04-17 | 电子科技大学 | Music list recommendation method and system based on long-term and short-term preference of user |
| US11636341B2 (en) | 2019-08-15 | 2023-04-25 | Advanced New Technologies Co., Ltd. | Processing sequential interaction data |
| CN110555469A (en) * | 2019-08-15 | 2019-12-10 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
| CN110598847A (en) * | 2019-08-15 | 2019-12-20 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
| US10936950B1 (en) | 2019-08-15 | 2021-03-02 | Advanced New Technologies Co., Ltd. | Processing sequential interaction data |
| US10970350B2 (en) | 2019-08-15 | 2021-04-06 | Advanced New Technologies Co., Ltd. | Method and apparatus for processing user interaction sequence data |
| US11250088B2 (en) | 2019-08-15 | 2022-02-15 | Advanced New Technologies Co., Ltd. | Method and apparatus for processing user interaction sequence data |
| CN110598847B (en) * | 2019-08-15 | 2020-08-28 | 阿里巴巴集团控股有限公司 | Method and device for processing interactive sequence data |
| CN110659742A (en) * | 2019-09-10 | 2020-01-07 | 阿里巴巴集团控股有限公司 | Method and device for acquiring sequence representation vector of user behavior sequence |
| CN110619081A (en) * | 2019-09-20 | 2019-12-27 | 苏州市职业大学 | News pushing method based on interactive graph neural network |
| CN110619081B (en) * | 2019-09-20 | 2022-05-17 | 苏州市职业大学 | News pushing method based on interactive graph neural network |
| CN110619082B (en) * | 2019-09-20 | 2022-05-17 | 苏州市职业大学 | Project recommendation method based on repeated search mechanism |
| CN110619082A (en) * | 2019-09-20 | 2019-12-27 | 苏州市职业大学 | Project recommendation method based on repeated search mechanism |
| CN110851705A (en) * | 2019-10-09 | 2020-02-28 | 天津大学 | Project-based collaborative storage recommendation method and recommendation device thereof |
| 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 |
| 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 |
| 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 |
| CN110929164A (en) * | 2019-12-09 | 2020-03-27 | 北京交通大学 | Interest point recommendation method based on user dynamic preference and attention mechanism |
| CN110929164B (en) * | 2019-12-09 | 2023-04-21 | 北京交通大学 | Point-of-interest recommendation method based on user dynamic preference and attention mechanism |
| CN113076469A (en) * | 2020-01-03 | 2021-07-06 | 阿里巴巴集团控股有限公司 | Model construction method, data recommendation method and device |
| CN113495991A (en) * | 2020-04-03 | 2021-10-12 | 北京沃东天骏信息技术有限公司 | Recommendation method and device |
| CN111582443B (en) * | 2020-04-22 | 2021-02-02 | 成都信息工程大学 | Recommendation method based on Mask mechanism and level attention mechanism |
| CN111582443A (en) * | 2020-04-22 | 2020-08-25 | 成都信息工程大学 | Recommendation method based on Mask mechanism and level attention mechanism |
| CN111552881B (en) * | 2020-05-09 | 2024-01-30 | 苏州市职业大学 | Sequence recommendation method based on hierarchical variation attention |
| CN111552881A (en) * | 2020-05-09 | 2020-08-18 | 苏州市职业大学 | Sequence recommendation method based on hierarchical variation attention |
| CN111753209B (en) * | 2020-07-02 | 2023-07-18 | 南京工业大学 | Sequence recommendation list generation method based on improved time sequence convolution network |
| CN111753209A (en) * | 2020-07-02 | 2020-10-09 | 南京工业大学 | Sequence recommendation list generation method based on improved time sequence convolutional network |
| CN112016000A (en) * | 2020-07-20 | 2020-12-01 | 桂林电子科技大学 | Movie group recommendation method and system based on convolution collaborative filtering |
| CN112256971A (en) * | 2020-10-29 | 2021-01-22 | 清华大学深圳国际研究生院 | Sequence recommendation method and computer-readable storage medium |
| CN112256971B (en) * | 2020-10-29 | 2023-06-20 | 清华大学深圳国际研究生院 | Sequence recommendation method and computer readable storage medium |
| CN113448245A (en) * | 2021-04-14 | 2021-09-28 | 华南师范大学 | Deep learning-based dissolved oxygen control method and system in sewage treatment process |
| CN113468434B (en) * | 2021-09-06 | 2021-12-24 | 北京搜狐新动力信息技术有限公司 | Resource recommendation method, device, readable medium and equipment |
| CN113468434A (en) * | 2021-09-06 | 2021-10-01 | 北京搜狐新动力信息技术有限公司 | Resource recommendation method, device, readable medium and equipment |
| CN113868542A (en) * | 2021-11-25 | 2021-12-31 | 平安科技（深圳）有限公司 | Attention model-based push data acquisition method, device, equipment and medium |
| CN114154071A (en) * | 2021-12-09 | 2022-03-08 | 电子科技大学 | Emotion time sequence recommendation method based on attention mechanism |
| CN114169869B (en) * | 2022-02-14 | 2022-06-07 | 北京大学 | Attention mechanism-based post recommendation method and device |
| CN114169869A (en) * | 2022-02-14 | 2022-03-11 | 北京大学 | Attention mechanism-based post recommendation method and device |
| CN115062230A (en) * | 2022-08-16 | 2022-09-16 | 深圳市海清视讯科技有限公司 | Data processing method, apparatus, device, medium, and program product |
| CN115545834A (en) * | 2022-10-10 | 2022-12-30 | 杭州电子科技大学 | Personalized service recommendation method based on graph neural network and metadata |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110060097A (en) | User behavior sequence of recommendation method based on attention mechanism and convolutional neural networks | |
| Dürr et al. | Probabilistic deep learning: With python, keras and tensorflow probability | |
| Jaafra et al. | Reinforcement learning for neural architecture search: A review | |
| Lu | Artificial intelligence: a survey on evolution, models, applications and future trends | |
| CN110119467B (en) | Project recommendation method, device, equipment and storage medium based on session | |
| Zellers et al. | From recognition to cognition: Visual commonsense reasoning | |
| Gulli et al. | Deep learning with Keras | |
| Guo et al. | Efficient hyperparameter optimization for convolution neural networks in deep learning: A distributed particle swarm optimization approach | |
| CN104217214B (en) | RGB D personage's Activity recognition methods based on configurable convolutional neural networks | |
| CN108564119A (en) | A kind of any attitude pedestrian Picture Generation Method | |
| CN109299396A (en) | Merge the convolutional neural networks collaborative filtering recommending method and system of attention model | |
| CN109754317B (en) | Comment-fused interpretable garment recommendation method, system, device and medium | |
| CN109902798A (en) | The training method and device of deep neural network | |
| CN107861938A (en) | A kind of POI official documents and correspondences generation method and device, electronic equipment | |
| CN107844784A (en) | Face identification method, device, computer equipment and readable storage medium storing program for executing | |
| CN108846314A (en) | A kind of food materials identification system and food materials discrimination method based on deep learning | |
| CN109657156A (en) | A kind of personalized recommendation method generating confrontation network based on circulation | |
| CN107563841A (en) | A kind of commending system decomposed that scored based on user | |
| CN109978836A (en) | User individual image esthetic evaluation method, system, medium and equipment based on meta learning | |
| CN112527966B (en) | Network text emotion analysis method based on Bi-GRU neural network and self-attention mechanism | |
| CN109101629A (en) | A kind of network representation method based on depth network structure and nodal community | |
| US20200272812A1 (en) | Human body part segmentation with real and synthetic images | |
| CN114693397A (en) | Multi-view multi-modal commodity recommendation method based on attention neural network | |
| CN106980831A (en) | Based on self-encoding encoder from affiliation recognition methods | |
| CN109447096A (en) | A kind of pan path prediction technique and device based on machine learning |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190726 |