CN117390300A - Construction method and device of multi-channel interactive learning interest point recommendation model - Google Patents
Construction method and device of multi-channel interactive learning interest point recommendation model Download PDFInfo
- Publication number
- CN117390300A CN117390300A CN202311301895.8A CN202311301895A CN117390300A CN 117390300 A CN117390300 A CN 117390300A CN 202311301895 A CN202311301895 A CN 202311301895A CN 117390300 A CN117390300 A CN 117390300A
- Authority
- CN
- China
- Prior art keywords
- user
- interest point
- interest
- features
- vector
- 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
- 230000002452 interceptive effect Effects 0.000 title claims abstract description 34
- 238000010276 construction Methods 0.000 title claims description 6
- 239000013598 vector Substances 0.000 claims abstract description 107
- 238000000034 method Methods 0.000 claims abstract description 56
- 239000011159 matrix material Substances 0.000 claims abstract description 49
- 230000007246 mechanism Effects 0.000 claims abstract description 34
- 238000013528 artificial neural network Methods 0.000 claims abstract description 24
- 230000003993 interaction Effects 0.000 claims abstract description 20
- 230000004931 aggregating effect Effects 0.000 claims abstract description 19
- 238000000354 decomposition reaction Methods 0.000 claims description 23
- 238000013507 mapping Methods 0.000 claims description 16
- 230000006870 function Effects 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 9
- 238000003860 storage Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 7
- 238000004422 calculation algorithm Methods 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 7
- 238000010606 normalization Methods 0.000 claims description 7
- 238000005295 random walk Methods 0.000 claims description 7
- 230000004927 fusion Effects 0.000 claims description 6
- 230000009466 transformation Effects 0.000 claims description 5
- 230000002776 aggregation Effects 0.000 claims description 3
- 238000004220 aggregation Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 claims description 2
- 238000012546 transfer Methods 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 5
- 238000013461 design Methods 0.000 description 4
- 230000003287 optical effect Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000001105 regulatory effect Effects 0.000 description 2
- 230000003068 static effect Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 239000004020 conductor Substances 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The disclosure provides a method and a device for constructing a multi-channel interactive learning interest point recommendation model, wherein the method comprises the following steps: aggregating the user potential representations and the interest point potential representations on the multiple element paths to obtain user multi-view semantic features and interest point multi-view semantic features, and respectively fusing the user multi-view semantic features and the interest point multi-view semantic features with corresponding internal features to obtain user comprehensive features and interest point comprehensive features; and utilizing a generalized matrix to decompose and learn low-order linear vectors of the user comprehensive features and the interest point comprehensive features, utilizing a multi-head self-attention mechanism deep neural network of convolution constraint to learn high-order nonlinear vectors of the user comprehensive features and the interest point comprehensive features, and fusing the low-order linear vectors and the high-order nonlinear vectors in a prediction layer to obtain the multi-channel interactive learning interest point recommendation model. The method and the device construct multi-view semantic features based on the meta-path, synchronously develop multi-channel learning of linear interaction and nonlinear interaction, and effectively capture a complex structure of user-interest point interaction.
Description
Technical Field
The present document relates to the field of computer technologies, and in particular, to a method and an apparatus for constructing a recommendation model of a multi-channel interactive learning interest point.
Background
The social network based on the location is a network for connecting the user and the interest point through the location information of the mobile device, the interest point recommendation is to utilize a large amount of location check-in data of the user in the social network, mining the life patterns and personal preferences of the user hidden behind the large-scale location data, and the social network recommendation is one of important services in the social network recommendation.
In the related technology, complex auxiliary information is often used for enriching the characteristic representation of the user and the interest point so as to relieve the high sparseness problem faced by the user-interest point sign-in matrix, but the auxiliary information built in the prior art has isomerism, so that the recommendation of the interest point is difficult to effectively utilize the auxiliary information; in addition, in interactive learning, a single matrix decomposition or deep neural network is generally used, so that a complex structure of interaction between a user and an interest point cannot be effectively captured, and an implicit feedback problem cannot be effectively processed.
By combining the analysis of the development status in the technical field, the prior technical proposal lacks auxiliary information construction integrating multi-view semantic features and distinguishing by using an attention mechanism, and a model for learning interactive features by using a multi-head self-attention mechanism deep neural network of generalized matrix decomposition and convolution constraint.
Disclosure of Invention
The invention aims to provide a method and a device for constructing a multi-channel interactive learning interest point recommendation model, and aims to solve the problems in the prior art.
According to a first aspect of an embodiment of the present invention, a method for constructing a multi-channel interactive learning interest point recommendation model is provided, including:
mapping the initial vectors of the user and the initial vectors of the interest points to an implicit vector space by using an embedding method to respectively obtain internal characteristics of the user and internal characteristics of the interest points;
aggregating the user neighbor pairs and the interest point neighbor pairs in the recommended meta-paths to obtain user initial isomorphic networks and interest point initial isomorphic networks corresponding to the meta-paths; adjusting the similarity between nodes in the initial isomorphic network of the user and the initial isomorphic network of the interest point through the adjustable parameter similarity to obtain a user weighted isomorphic network and an interest point weighted isomorphic network; projecting the user weighted isomorphic network and the interest point weighted isomorphic network to a low-dimensional space by using a random walk algorithm to obtain a user potential representation and an interest point potential representation;
respectively gathering user potential representations and interest point potential representations on a plurality of element paths by using an attention mechanism to obtain user multi-view semantic features and interest point multi-view semantic features;
fusing the internal features of the user with the multi-view semantic features of the user to obtain comprehensive features of the user, and fusing the internal features of the interest points with the multi-view semantic features of the interest points to obtain comprehensive features of the interest points;
and inputting the comprehensive characteristics of the user and the comprehensive characteristics of the interest points into a generalized matrix for decomposition to obtain low-order linear vectors, inputting the comprehensive characteristics of the user and the comprehensive characteristics of the interest points into a convolution-constrained multi-head self-attention mechanism deep neural network to obtain high-order nonlinear vectors, and fusing the low-order linear vectors and the high-order nonlinear vectors in a prediction layer to obtain the multi-channel interactive learning interest point recommendation model.
According to a second aspect of the embodiment of the present invention, there is provided a device for constructing a multi-channel interactive learning interest point recommendation model, including:
the embedding module is used for mapping the initial vectors of the user and the initial vectors of the interest points to an implicit vector space by using an embedding method to respectively obtain internal characteristics of the user and internal characteristics of the interest points;
the potential representation generation module is used for gathering the user neighbor pairs and the interest point neighbor pairs in the recommended meta-paths to obtain user initial isomorphic networks and interest point initial isomorphic networks corresponding to the multiple meta-paths; adjusting the similarity between nodes in the user initial network and the interest point initial isomorphic network through the adjustable parameter similarity to obtain a user weighted isomorphic network and an interest point weighted isomorphic network; projecting the user weighted isomorphic network and the interest point weighted isomorphic network to a low-dimensional space by using a random walk algorithm to obtain a user potential representation and an interest point potential representation;
the potential representation aggregation module is used for respectively aggregating the potential representations of the user and the potential representations of the interest points on a plurality of element paths by using an attention mechanism to obtain multi-view semantic features of the user and multi-view semantic features of the interest points;
the comprehensive feature fusion module is used for fusing the internal features of the user with the multi-view semantic features of the user to obtain comprehensive features of the user, and fusing the internal features of the interest points with the multi-view semantic features of the interest points to obtain comprehensive features of the interest points;
the multi-channel interaction construction module is used for inputting the comprehensive characteristics of the user and the comprehensive characteristics of the interest points into a generalized matrix for decomposition to obtain a low-order linear vector, inputting the comprehensive characteristics of the user and the comprehensive characteristics of the interest points into a convolution constraint multi-head self-attention mechanism deep neural network to obtain a high-order nonlinear vector, and fusing the low-order linear vector and the high-order nonlinear vector in a prediction layer to obtain the multi-channel interaction learning interest point recommendation model.
According to a third aspect of an embodiment of the present invention, there is provided an electronic apparatus including: a memory, a processor, and a computer program stored on the memory and executable on the processor, which when executed by the processor, performs the steps of a method of constructing a multi-channel interactive learning point of interest recommendation model as provided in the first aspect of the present disclosure.
According to a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium having stored thereon a program for implementing information transfer, which when executed by a processor, implements the steps of the method for constructing a multi-channel interactive learning interest point recommendation model provided in the first aspect of the present disclosure.
The technical scheme provided by the embodiment of the invention has the following beneficial effects: the potential representation corresponding to the weighted isomorphic network formed by the isomorphic networks on a plurality of element paths is gathered by using an attention mechanism, so that the characteristic representation of users and interest points is enriched, and the heterogeneous characteristics of data sparseness and auxiliary information are effectively relieved; the multi-head self-attention mechanism deep neural network utilizing generalized matrix decomposition and convolution constraint learns the characteristic interaction between the user and the interest point, thereby effectively relieving the implicit feedback problem.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.
Drawings
For a clearer description of one or more embodiments of the present description or of the solutions of the prior art, the drawings that are necessary for the description of the embodiments or of the prior art will be briefly described, it being apparent that the drawings in the description that follow are only some of the embodiments described in the description, from which, for a person skilled in the art, other drawings can be obtained without inventive faculty.
FIG. 1 is a flowchart of a method for constructing a multi-channel interactive learning interest point recommendation model according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a point of interest recommendation model design framework in accordance with an embodiment of the present invention;
FIG. 3 is a schematic diagram of a device for constructing a multi-channel interactive learning interest point recommendation model according to an embodiment of the present invention;
fig. 4 is a schematic diagram of an electronic device according to an embodiment of the invention.
Detailed Description
In order to enable a person skilled in the art to better understand the technical solutions in one or more embodiments of the present specification, the technical solutions in one or more embodiments of the present specification will be clearly and completely described below with reference to the drawings in one or more embodiments of the present specification, and it is obvious that the described embodiments are only some embodiments of the present specification, not all embodiments. All other embodiments, which can be made by one or more embodiments of the present disclosure without inventive faculty, are intended to be within the scope of the present disclosure.
Method embodiment
According to an embodiment of the present invention, a method for constructing a multi-channel interactive learning interest point recommendation model is provided, and fig. 1 is a flowchart of a method for constructing a multi-channel interactive learning interest point recommendation model according to an embodiment of the present invention, as shown in fig. 1, where the method for constructing a multi-channel interactive learning interest point recommendation model according to an embodiment of the present invention specifically includes:
in step S110, the user initial vector and the interest point initial vector are mapped to an implicit vector space using an embedding method, so as to obtain a user internal feature and an interest point internal feature, respectively. The method specifically comprises the following steps:
performing one-hot coding on the user features and the interest point features, mapping the user initial vector and the interest point initial vector to implicit space vectors respectively by using an embedding method, namely converting an original sparse high-dimensional representation into a dense low-dimensional representation by potential factor matrix adjustment to obtain user internal features and interest point internal features, mapping the user initial vector by using a formula 1, and mapping the interest point initial vector by using a formula 2:
in the method, in the process of the invention,representing the internal characteristics of the user->Representing the internal characteristics of the interest points; />And->Representation adopts embedding method mapping; u (U) u Representing the user initial vector, P p Representing an initial vector of points of interest; d (D) T Representing a matrix of user potential factors, Q T Representing a matrix of point of interest latent factors.
In step S120, user neighbor pairs and interest point neighbor pairs in the recommended meta-paths are aggregated to obtain a user initial isomorphic network and an interest point initial isomorphic network corresponding to the multiple meta-paths; adjusting the similarity between nodes in the initial isomorphic network of the user and the initial isomorphic network of the interest point through the adjustable parameter similarity to obtain a user weighted isomorphic network and an interest point weighted isomorphic network; and projecting the user weighted isomorphic network and the interest point weighted isomorphic network to a low-dimensional space by using a random walk algorithm to obtain the user potential representation and the interest point potential representation. The method specifically comprises the following steps:
acquiring a node set related to a user through a meta-path, if the node set related to the user contains another user, forming the two users into a user neighbor pair, and aggregating all the user neighbor pairs in the meta-path to obtain a user initial isomorphic network of the meta-path; aggregating user neighbor pairs on different element paths to obtain user initial isomorphic networks corresponding to a plurality of element paths;
acquiring a node set related to a certain interest point through a certain element path, if the node set related to the interest point contains another interest point, forming interest point neighbor pairs by the two interest points, and aggregating all interest point neighbor pairs in the element path to obtain an interest point initial isomorphic network of the element path; and gathering interest point neighbor pairs on different element paths to obtain interest point initial isomorphic networks corresponding to a plurality of element paths.
To aggregate element paths ζ m For example, if the secondary path ζ is m Is obtained from user u i Related node setAssume that there is another node +>Then a meta-path ζ is formed m A guided pair of neighbor pairs ζ m <u i ,u j >Aggregating element paths ζ m Neighbor pairs of all the nodes on the network get the meta-path xi m Corresponding user initial isomorphic network, if any two neighbor pairs have the same node, forming a network structure by using the node as a connecting node, and generating a meta-path xi by using a formula 3 m Corresponding user initial homogeneous network->
Where g () represents the integrated operation of aggregating neighbor pairs, u s Representing the originating neighbor node, u t Indicating that the neighbor node is to be terminated,representing meta-path xi m A set of all user nodes. The types of the meta-paths used for recommendation are more, such as that two users visit the same interest point, two users visit the interest point of the same area, and the like, and the corresponding interest point initial isomorphic networks are obtained on different meta-paths.
The similarity between nodes in the initial isomorphic network is adjusted through the adjustable parameter similarity, so as to distinguish the affinity and the sparsity between the nodes, wherein the adjustable parameter similarity is Jaccard similarity using the self-defined adjustable parameter, and the similarity between the nodes is adjusted by using a formula 4:
in the method, in the process of the invention,representing node u i And node u j Similarity between->Representing meta-path xi m Go up u i And u j The number of paths between->And->Respectively represent meta-paths xi m Go up u i And u j Neighbor number of->User-defined adjustable parameters for representing influence of balanced path numbers on similarity are obtained after the similarity is adjusted, a user weighted isomorphic network and interest point weighted isomorphic networks are obtained, and the user weighted isomorphic networks corresponding to a plurality of element paths are represented as +.>Where ζ represents the set of meta-paths.
Using random walk algorithm to project the user weighted isomorphic network and interest point weighted isomorphic network to low-dimensional space to obtain more efficient and interpretable feature representation, taking the user weighted isomorphic network as an example, forEach edge (u) i ,u j ) Node u is defined using equation 5 i And u j Joint probabilities between nodes, the empirical joint probability between nodes is defined using equation 6:
wherein P is r (u i ,u j ) The joint probability is represented as a function of the joint probability,and->Respectively represent u i And u j Is represented by hidden vectors of->Representing empirical joint probability, ++>Representing edge (u) i ,u j ) Weight of->Representation->The set of middle edges selects the KL divergence of two probability distributions as an optimization target, and the form is shown in the formula 7:
wherein O represents an optimization target, and the optimization target is minimized to obtainLow-dimensional vector representation of upper user nodeWherein (1)>Representing node u in meta-path ζ m User potential representation on->Representation->A set of user nodes.
In step S130, the user potential representations and the interest point potential representations on the multiple meta-paths are respectively aggregated by using the attention mechanism, so as to obtain the multi-view semantic features of the user and the multi-view semantic features of the interest point. The method specifically comprises the following steps:
the potential representations of nodes on the meta-paths only contain neighbor information and semantic information inside the corresponding meta-paths, the potential representations on all meta-paths need to be integrated, different users prefer different meta-paths, and therefore the differences need to be measured.
Setting user computing weights of different element paths for a user by using an attention mechanism, and gathering potential representations of the user on the different element paths through the user computing weights to obtain multi-view semantic features of the user corresponding to the user; and setting interest point calculation weights of different element paths for the interest points by using an attention mechanism, and gathering potential representations of the interest points on the different element paths through the interest point calculation weights to obtain multi-view semantic features of the interest points corresponding to the interest points.
Taking user u and point of interest p as examples, the user potential on the meta-path is represented asInterest points of interest points on the meta path are potentially denoted +.>User u is assigned the weights of the different meta-paths using equation 8 and point of interest p is assigned the weights of the different meta-paths using equation 9:
in the method, in the process of the invention,representing meta-path xi m Final weight for user u, +.>Representing meta-path xi n Final weight for point of interest p; />And->Weight representing the attention network, +.>And->Representing the bias, ζ, of the network (U) And xi (P) Representing the set of u and p-element paths, respectively, relu () represents the activation function and exp represents the e-based exponential function. According to the attention weight of each element path, obtaining the multi-view semantic features of the user and the multi-view semantic features of the interest points, and calculating the multi-view semantic features h of the user u by using a formula 10 u Calculating a point of interest p final potential representation h of the point of interest using equation 11 p ：
In step S140, the user internal feature and the user multi-view semantic feature are fused to obtain a user comprehensive feature, and the interest point internal feature and the interest point multi-view semantic feature are fused to obtain an interest point comprehensive feature. The method specifically comprises the following steps:
for user u, the user internal features obtained in step S110 are formulated by equation 12And the user multi-view semantic feature h obtained in the step S130 u Fusion is carried out to obtain the comprehensive characteristics H of the user U The method comprises the steps of carrying out a first treatment on the surface of the For the interest point p, the interest point obtained in step S110 is internally calculated by equation 13Characteristics->And the interest point multi-view semantic feature h obtained in the step S130 p Fusion is carried out to obtain the comprehensive feature H of the interest point P ：
In step S150, the user comprehensive features and the interest point comprehensive features are input into a generalized matrix for decomposition to obtain a low-order linear vector, the user comprehensive features and the interest point comprehensive features are input into a convolution constraint multi-head self-attention mechanism deep neural network to obtain a high-order nonlinear vector, and the low-order linear vector and the high-order nonlinear vector are fused in a prediction layer to obtain a multi-channel interactive learning interest point recommendation model. The method specifically comprises the following steps:
the traditional matrix decomposition estimates the interaction between the features as the inner product of the potential feature representation, the generalized matrix decomposition estimates the interaction as the element product of the potential feature representation, the user comprehensive features and the interest point comprehensive features are input into the generalized matrix decomposition, the Hadamard product interaction between the user comprehensive features and the interest point comprehensive features is calculated through the generalized matrix decomposition, the low-order linear vector through the element product is obtained, and the generalized matrix decomposition mapping is carried out through a formula 14:
in the method, in the process of the invention,representing a low order linear vector, phi representing the mapping function of the Hadamard product interaction, +.>Representing element-wise multiplication between vectors.
Inputting the user comprehensive features and the interest point comprehensive features into a left nonlinear layer formed by a multi-layer perceptron, and outputting left vectors through the left nonlinear layer, wherein the multi-layer perceptron comprises a plurality of hidden layers, the hidden layer design mode is halved step by step according to the dimension, one Dropout layer is added behind each hidden layer for preventing overfitting and improving the generalization capability of the model, the user comprehensive features and the interest point comprehensive features are combined into features Z through a formula 15, and the definition of the multi-layer perceptron MLP is shown as a formula 16:
wherein, psi represents vector concatenation, y MLP Representing the left nonlinear layer output vector, W X Representing the weight matrix of the layer X perceptron,represents W X Reverse of (b) X Representing the offset vector, alpha, of the layer X perceptron X Representing an activation function of the X-layer perceptron; and the relu () is used as an activation function of the multi-layer perceptron, has unsaturated characteristics and has good adaptability to sparse functions.
The multi-head self-attention mechanism can learn nonlinear relations between users and interest points, including high-order interaction modes and complex dependency relations, different relations can be focused through a plurality of attention heads, and the comprehensive characteristics of the users and the comprehensive characteristics of the interest points are input into a right nonlinear layer formed by a multi-head attention module and a feedforward neural network; in the multi-head attention module, a linear mapping matrix W is used based on equation 17 q 、W k And W is v Decomposing feature groups into vectorsq, k and v:
q and k are calculated by equation 18 T The scaling dot product attention is adopted to calculate the weight of the two, the weight is normalized by a softmax function, a local receptive field of convolution operation is introduced to obtain an attention output matrix, and the normalized matrix is subjected to dot multiplication by convolution constraint v to obtain the attention output matrix: the convolution operation is calculated using equation 19:
f(v)＝relu(v☉W c ) Equation 19;
wherein Attention (q, k, v) represents generating Attention output matrix corresponding to vectors q, k and v, softmax () represents softmax normalization, i.e. inverting k of vector q and vector k T The normalization processing is carried out after the calculation of the scaling dot product,dimension d representing k vector k F (v) represents the square root of the convolution operation, W c Parameters representing the convolution kernel, relu () represents the activation function, +..
Calculating multiple attention scores through a formula 20, splicing all scores through a formula 21 and considering attention head weights to obtain vectors output by the multiple attention modules:
head i =attention (q, k, v) formula 20;
Y＝MHA(Z)＝concat(head 1 ,head 2 ......head h )W Y equation 21;
wherein head i Weight vector representing i-th attention header output, concat () represents concatenation, W Y Representing a weight matrix, MHA () represents a multi-head attention map, Y represents a vector output by a multi-head attention module, and the vector is represented by a formula22 non-linear transformation is performed on the vector output by the multi-head attention module by using a feedforward neural network, and after transformation in the embodiment, normalization, residual connection and Dropout layers are respectively added for the multi-head attention module and the feedforward neural network module by using a formula 23 and a formula 24 to prevent unstable training:
FFN(Y)＝relu(YW 1 +b 1 )W 2 +b 2 equation 22;
y MHA =ln (Dropout (MHA (Z))+z formula 23;
y FFN =ln (Dropout (FFN (Y))+y formula 24;
wherein FFN () represents feed forward mapping, LN () represents nonlinear transformation, vector y MHA Sum vector y FFN Representing the result of multi-head attention and the result of feedforward neural network mapping, respectively, wherein vector y FFN To the right vector; fusing the left side vector and the right side vector through a formula 25 to obtain a high-order nonlinear vector
The multi-headed self-attention mechanism deep neural network with generalized matrix decomposition and convolution constraint respectively learns to embed independently, and combines two models by connecting one hidden layer, as shown in formula 26:
wherein sigma represents a Sigmoid () activation function, eta represents a connection weight, and the relation between the multi-head self-attention mechanism deep neural network for regulating generalized matrix decomposition and convolution constraint is regulated by outputting vectorsOutputting probability of recommending interest points for user。
The method further comprises the steps of:
in step S160, a multi-channel interactive learning interest point recommendation model is used to recommend interest points for the user, that is, calculate the probability of recommending different interest points for the user, and recommend the interest point with the highest probability for the user.
The above technical solutions of the embodiments of the present invention are illustrated with reference to the following drawings.
FIG. 2 is a schematic diagram of a design framework of a point of interest recommendation model according to an embodiment of the present invention, as shown in FIG. 2, illustrating a complete design scheme for constructing a point of interest recommendation model, including mapping implicit space using an embedding method; constructing a isomorphic network; aggregating user potential representations and interest point potential representations corresponding to weighted isomorphic networks formed by isomorphic networks on different element paths using an attention mechanism; and obtaining user comprehensive features and interest point comprehensive features through feature fusion, respectively performing generalized matrix decomposition and multi-head self-attention mechanism deep neural network learning of convolution constraint on the fused features, and fusing prediction result vectors in a prediction layer.
In summary, aiming at the problems existing in the current situation, the method for constructing the multi-channel interactive learning interest point recommendation model uses Jaccard similarity with the user-defined adjustable parameters to adjust the similarity among nodes in the initial isomorphic network, so that the influence of the path number on the similarity can be balanced; the potential representations corresponding to the weighted isomorphic networks on a plurality of element paths are gathered by using an attention mechanism, so that the potential representations of users and interest points are enriched, and the heterogeneous characteristics of data sparseness and auxiliary information are effectively relieved; the multi-head self-attention mechanism deep neural network utilizing generalized matrix decomposition and convolution constraint is utilized to learn the characteristic interaction between the user and the interest point, so that the implicit feedback problem is effectively relieved; the multi-head self-attention deep learning network distinguishes the correlation among different characteristic domains, and introduces a local receptive field of convolution operation to obtain an attention output matrix, so that excessive redundancy of attention units is effectively prevented, and the generalization capability of a model is improved.
Device embodiment
According to an embodiment of the present invention, a device for constructing a multi-channel interactive learning interest point recommendation model is provided, and fig. 3 is a schematic diagram of the device for constructing a multi-channel interactive learning interest point recommendation model according to the embodiment of the present invention, as shown in fig. 3, where the device for constructing a multi-channel interactive learning interest point recommendation model according to the embodiment of the present invention specifically includes:
the embedding module 30 is configured to map the user initial vector and the interest point initial vector to an implicit vector space by using an embedding method, so as to obtain the user internal feature and the interest point internal feature respectively.
A potential representation generating module 32, configured to aggregate the user neighbor pairs and the interest point neighbor pairs in the meta-paths for recommendation, so as to obtain a user initial isomorphic network and an interest point initial isomorphic network corresponding to the multiple meta-paths; adjusting the similarity between nodes in the initial isomorphic network of the user and the initial isomorphic network of the interest point through the adjustable parameter similarity to obtain a user weighted isomorphic network and an interest point weighted isomorphic network; and projecting the user weighted isomorphic network and the interest point weighted isomorphic network to a low-dimensional space by using a random walk algorithm to obtain the user potential representation and the interest point potential representation. The method is particularly used for:
acquiring a node set related to a user through a meta-path, if the node set related to the user contains another user, forming the two users into a user neighbor pair, and aggregating all the user neighbor pairs in the meta-path to obtain a user initial isomorphic network of the meta-path; aggregating neighbor pairs on different element paths to obtain user initial isomorphic networks corresponding to a plurality of element paths;
acquiring a node set related to a certain interest point through a certain element path, if the node set related to the interest point contains another interest point, forming interest point neighbor pairs by the two interest points, and aggregating all interest point neighbor pairs in the element path to obtain an interest point initial isomorphic network of the element path; and gathering interest point neighbor pairs on different element paths to obtain interest point initial isomorphic networks corresponding to a plurality of element paths.
The similarity between nodes in the initial isomorphic network is adjusted through the adjustable parameter similarity based on the formula 1, wherein the adjustable parameter similarity is Jaccard similarity using the self-defined adjustable parameters:
wherein,representing node u i And node u j Similarity between->Representing meta-path xi m Go up u i And u j The number of paths between->And->Respectively represent meta-paths xi m Go up u i And u j Neighbor number of->A custom adjustable parameter representing the effect of the number of balanced paths on similarity.
The potential representation aggregation module 34 is configured to aggregate the user potential representations and the interest point potential representations on the multiple meta-paths by using an attention mechanism, so as to obtain a multi-view semantic feature of the user and a multi-view semantic feature of the interest point. The method is particularly used for:
setting user computing weights of different element paths for a user by using an attention mechanism, and gathering potential representations of the user on the different element paths through the user computing weights to obtain multi-view semantic features of the user corresponding to the user;
and setting interest point calculation weights of different element paths for the interest points by using an attention mechanism, and gathering potential representations of the interest points on the different element paths through the interest point calculation weights to obtain multi-view semantic features of the interest points corresponding to the interest points.
The comprehensive feature fusion module 36 is configured to fuse the internal features of the user with the multi-view semantic features of the user to obtain comprehensive features of the user, and fuse the internal features of the interest points with the multi-view semantic features of the interest points to obtain comprehensive features of the interest points.
The multi-channel interaction construction module 38 is configured to input the user integrated feature and the interest point integrated feature into a generalized matrix, decompose the generalized matrix to obtain a low-order linear vector, input the user integrated feature and the interest point integrated feature into a multi-head self-attention mechanism deep neural network with convolution constraint to obtain a high-order nonlinear vector, and fuse the low-order linear vector and the high-order nonlinear vector in a prediction layer to obtain the multi-channel interaction learning interest point recommendation model. The method is particularly used for:
inputting the comprehensive characteristics of the user and the comprehensive characteristics of the interest points into generalized matrix decomposition, and calculating Hadamard product interaction between the comprehensive characteristics of the user and the comprehensive characteristics of the interest points through the generalized matrix decomposition to obtain a low-order linear vector passing through element products.
Inputting the comprehensive characteristics of the user and the comprehensive characteristics of the interest points into a left nonlinear layer formed by a multi-layer perceptron, and outputting a left vector through the left nonlinear layer, wherein the multi-layer perceptron comprises a plurality of hidden layers, and a Dropout layer is added behind each hidden layer;
inputting the comprehensive characteristics of the user and the comprehensive characteristics of the interest points into a right nonlinear layer consisting of a multi-head attention module and a feedforward neural network; decomposing the feature group into vectors q, k and v by using a linear mapping matrix in a multi-head attention module, introducing a local receptive field of convolution operation to obtain an attention output matrix, obtaining a score of each attention head by the attention output matrix, splicing all the scores and considering the weight of the attention heads to obtain a vector output by the multi-head attention module, generating the attention output matrix by using a formula 2, and calculating the convolution operation by using a formula 3:
f(v)＝relu(v☉W c ) Equation 3;
wherein, attention is%q, k, v) represents generating the attention output matrix corresponding to the vectors q, k and v, and softmax () represents softmax normalization, i.e., the inverse k of the vector q and the vector k T The normalization processing is carried out after the calculation of the scaling dot product,dimension d representing k vector k F (v) represents the square root of the convolution operation, W c Parameters representing the convolution kernel, relu () represents the activation function, +..
Nonlinear transformation is carried out on vectors output by the multi-head attention module by using a feedforward neural network, so that right side vectors are obtained; and fusing the left side vector and the right side vector to obtain a high-order nonlinear vector.
The method further comprises the steps of:
a recommendation module 310 for recommending points of interest to a user using a multi-channel interactive learning point of interest recommendation model.
In summary, aiming at the problems existing in the current situation, the device for constructing the multi-channel interactive learning interest point recommendation model disclosed by the invention uses Jaccard similarity with self-defined adjustable parameters to adjust the similarity among nodes in an initial isomorphic network, so that the influence of the number of paths on the similarity can be balanced; the potential representations corresponding to the weighted isomorphic networks on a plurality of element paths are gathered by using an attention mechanism, so that the potential representations of users and interest points are enriched, and the heterogeneous characteristics of data sparseness and auxiliary information are effectively relieved; the multi-head self-attention mechanism deep neural network utilizing generalized matrix decomposition and convolution constraint is utilized to learn the characteristic interaction between the user and the interest point, so that the implicit feedback problem is effectively relieved; the multi-head self-attention deep learning network distinguishes the correlation among different characteristic domains, and introduces a local receptive field of convolution operation to obtain an attention output matrix, so that excessive redundancy of attention units is effectively prevented, and the generalization capability of a model is improved.
Electronic device embodiment
Fig. 4 is a schematic diagram of an electronic device according to an embodiment of the invention. Electronic device 400 may include at least one processor 410 and memory 420. Processor 410 may execute instructions stored in memory 420. The processor 410 is communicatively coupled to the memory 420 via a data bus. In addition to memory 420, processor 410 may be communicatively coupled with input device 430, output device 440, and communication device 450 via a data bus.
The processor 410 may be any conventional processor, such as a commercially available CPU. The processor may also include, for example, an image processor (Graphic Process Unit, GPU), a field programmable gate array (Field Programmable Gate Array, FPGA), a System On Chip (SOC), an application specific integrated Chip (Application Specific Integrated Circuit, ASIC), or a combination thereof.
The memory 420 may be implemented by any type or combination of volatile or nonvolatile memory devices such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disk.
In the embodiment of the present disclosure, the memory 420 stores executable instructions, and the processor 410 may read the executable instructions from the memory 420 and execute the instructions to implement all or part of the steps of the method for constructing the multi-channel interactive learning interest point recommendation model in any of the above exemplary embodiments.
Computer-readable storage medium embodiments
In addition to the methods and apparatus described above, exemplary embodiments of the present disclosure may also be a computer program product or a computer-readable storage medium storing the computer program product, the computer program product including computer program instructions executable by a processor to implement all or part of the steps described in the method of constructing a multi-channel interactive learning point of interest recommendation model in any of the exemplary embodiments described above.
The computer program product may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages, and scripting languages (e.g., python). The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device, partly on a remote computing device, or entirely on the remote computing device or server.
A computer readable storage medium may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may include, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples of the readable storage medium include: a Static Random Access Memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic or optical disk, or any suitable combination of the foregoing having one or more electrical conductors.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.
Claims (10)
1. The method for constructing the multi-channel interactive learning interest point recommendation model is characterized by comprising the following steps of:
mapping the initial vectors of the user and the initial vectors of the interest points to an implicit vector space by using an embedding method to respectively obtain internal characteristics of the user and internal characteristics of the interest points;
aggregating the user neighbor pairs and the interest point neighbor pairs in the recommended meta-paths to obtain user initial isomorphic networks and interest point initial isomorphic networks corresponding to the meta-paths; adjusting the similarity between nodes in the initial isomorphic network of the user and the initial isomorphic network of the interest point through the adjustable parameter similarity to obtain a user weighted isomorphic network and an interest point weighted isomorphic network; projecting the user weighted isomorphic network and the interest point weighted isomorphic network to a low-dimensional space by using a random walk algorithm to obtain a user potential representation and an interest point potential representation;
respectively gathering the user potential representation and the interest point potential representation on a plurality of meta-paths by using an attention mechanism to obtain user multi-view semantic features and interest point multi-view semantic features;
fusing the internal features of the user with the multi-view semantic features of the user to obtain comprehensive features of the user, and fusing the internal features of the interest points with the multi-view semantic features of the interest points to obtain comprehensive features of the interest points;
and inputting the user comprehensive features and the interest point comprehensive features into a generalized matrix for decomposition to obtain a low-order linear vector, inputting the user comprehensive features and the interest point comprehensive features into a convolution constraint multi-head self-attention mechanism deep neural network to obtain a high-order nonlinear vector, and fusing the low-order linear vector and the high-order nonlinear vector in a prediction layer to obtain a multi-channel interactive learning interest point recommendation model.
2. The method according to claim 1, wherein the method further comprises:
and recommending the interest points for the user by using the multi-channel interactive learning interest point recommendation model.
3. The method of claim 1, wherein the aggregating the pairs of user neighbors and pairs of interest neighbors in the recommended meta-paths to obtain the user initial homogeneous network and the interest initial homogeneous network corresponding to the plurality of meta-paths specifically comprises:
acquiring a node set related to a user through a meta-path, if the node set related to the user contains another user, forming a user neighbor pair by the two users, and aggregating all the user neighbor pairs in the meta-path to obtain a user initial isomorphic network of the meta-path; aggregating user neighbor pairs on different element paths to obtain user initial isomorphic networks corresponding to a plurality of element paths;
acquiring a node set related to a certain interest point through a certain element path, if the node set related to the interest point contains another interest point, forming interest point neighbor pairs by the two interest points, and aggregating all the interest point neighbor pairs in the element path to obtain an interest point initial isomorphic network of the element path; and gathering interest point neighbor pairs on different element paths to obtain interest point initial isomorphic networks corresponding to a plurality of element paths.
4. The method of claim 1, wherein the adjusting the similarity between the nodes in the user-initiated homogeneous network and the point-of-interest-initiated homogeneous network by the adjustable parameter similarity to obtain the user-weighted homogeneous network and the point-of-interest-weighted homogeneous network specifically comprises:
the similarity between nodes in the initial isomorphic network is adjusted through the adjustable parameter similarity based on the formula 1, wherein the adjustable parameter similarity is Jaccard similarity using the self-defined adjustable parameters:
wherein,representing node u i And node u j Similarity between->Representing meta-path xi m Go up u i And u j The number of paths between->And->Respectively represent meta-paths xi m Go up u i And u j Neighbor number of->A custom adjustable parameter representing the effect of the number of balanced paths on similarity.
5. The method according to claim 1, wherein the aggregating the user potential representations and the interest point potential representations on the plurality of meta-paths by using an attention mechanism to obtain the user multi-view semantic feature and the interest point multi-view semantic feature comprises:
setting user computing weights of different element paths for a user by using an attention mechanism, and gathering potential representations of the user on the different element paths through the user computing weights to obtain multi-view semantic features of the user corresponding to the user;
and setting interest point calculation weights of different element paths for the interest points by using an attention mechanism, and gathering potential representations of the interest points on the different element paths through the interest point calculation weights to obtain interest point multi-view semantic features corresponding to the interest points.
6. The method of claim 1, wherein said inputting the user integrated feature and the interest point integrated feature into a generalized matrix decomposition, obtaining a low-order linear vector, specifically comprises:
and inputting the user comprehensive features and the interest point comprehensive features into generalized matrix decomposition, and calculating Hadamard product interaction between the user comprehensive features and the interest point comprehensive features through the generalized matrix decomposition to obtain low-order linear vectors passing through element products.
7. The method of claim 1, wherein the inputting the user-integrated feature and the point-of-interest integrated feature into the convolution constrained multi-head self-attention mechanism deep neural network, the obtaining a high-order nonlinear vector specifically comprises:
inputting the user comprehensive characteristics and the interest point comprehensive characteristics into a left nonlinear layer formed by a multi-layer perceptron, and outputting a left vector through the left nonlinear layer, wherein the multi-layer perceptron comprises a plurality of hidden layers, and a Dropout layer is added behind each hidden layer;
inputting the user comprehensive characteristics and the interest point comprehensive characteristics into a right nonlinear layer consisting of a multi-head attention module and a feedforward neural network; decomposing a feature group into vectors q, k and v by using a linear mapping matrix in the multi-head attention module, introducing a local receptive field of convolution operation to obtain an attention output matrix, obtaining a score of each attention head by using the attention output matrix, splicing all scores and considering the attention head weight to obtain a vector output by the multi-head attention module, generating the attention output matrix by using a formula 2, and calculating the convolution operation by using a formula 3:
f(v)＝relu(v⊙W c ) Equation 3;
wherein Attention (q, k, v) represents generating Attention output matrix corresponding to vectors q, k and v, softmax () represents softmax normalization, i.e. inverting k of vector q and vector k T The normalization processing is carried out after the calculation of the scaling dot product,dimension d representing k vector k F (v) represents the square root of the convolution operation, W c Parameters representing the convolution kernel, relu () representing the activation function, +.;
the feedforward neural network is used for carrying out nonlinear transformation on vectors output by the multi-head attention module to obtain right side direction; and fusing the left side vector and the right side vector to obtain a high-order nonlinear vector.
8. The device for constructing the multi-channel interactive learning interest point recommendation model is characterized by comprising the following components:
the embedding module is used for mapping the initial vectors of the user and the initial vectors of the interest points to an implicit vector space by using an embedding method to respectively obtain internal characteristics of the user and internal characteristics of the interest points;
the potential representation generation module is used for gathering the user neighbor pairs and the interest point neighbor pairs in the recommended meta-paths to obtain user initial isomorphic networks and interest point initial isomorphic networks corresponding to the multiple meta-paths; adjusting the similarity between nodes in the initial isomorphic network of the user and the initial isomorphic network of the interest point through the adjustable parameter similarity to obtain a user weighted isomorphic network and an interest point weighted isomorphic network; projecting the user weighted isomorphic network and the interest point weighted isomorphic network to a low-dimensional space by using a random walk algorithm to obtain a user potential representation and an interest point potential representation;
the potential representation aggregation module is used for respectively aggregating the user potential representations and the interest point potential representations on a plurality of meta-paths by using an attention mechanism to obtain user multi-view semantic features and interest point multi-view semantic features;
the comprehensive feature fusion module is used for fusing the internal features of the user with the multi-view semantic features of the user to obtain comprehensive features of the user, and fusing the internal features of the interest points with the multi-view semantic features of the interest points to obtain comprehensive features of the interest points;
the multi-channel interaction construction module is used for inputting the user comprehensive features and the interest point comprehensive features into a generalized matrix for decomposition to obtain a low-order linear vector, inputting the user comprehensive features and the interest point comprehensive features into a convolution constraint multi-head self-attention mechanism deep neural network to obtain a high-order nonlinear vector, and fusing the low-order linear vector and the high-order nonlinear vector in a prediction layer to obtain a multi-channel interaction learning interest point recommendation model.
9. An electronic device, comprising: memory, a processor and a computer program stored on the memory and executable on the processor, which when executed by the processor, implements the steps of the method of constructing a multi-channel interactive learning point of interest recommendation model as claimed in any one of claims 1 to 7.
10. A computer-readable storage medium, wherein a program for implementing information transfer is stored on the computer-readable storage medium, and the program when executed by a processor implements the steps of the method for constructing a multi-channel interactive learning interest point recommendation model according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311301895.8A CN117390300B (en) | 2023-10-09 | Construction method and device of multi-channel interactive learning interest point recommendation model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311301895.8A CN117390300B (en) | 2023-10-09 | Construction method and device of multi-channel interactive learning interest point recommendation model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117390300A true CN117390300A (en) | 2024-01-12 |
| CN117390300B CN117390300B (en) | 2024-07-02 |
Family
ID=
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160157057A1 (en) * | 2014-04-22 | 2016-06-02 | Sk Planet Co., Ltd. | Device for recommending location in building by using fingerprint of access point, and method using same |
| CN112069415A (en) * | 2020-08-13 | 2020-12-11 | 中国海洋大学 | Interest point recommendation method based on heterogeneous attribute network characterization learning |
| US20210248461A1 (en) * | 2020-02-11 | 2021-08-12 | Nec Laboratories America, Inc. | Graph enhanced attention network for explainable poi recommendation |
| CN115310005A (en) * | 2022-08-15 | 2022-11-08 | 中国人民解放军战略支援部队信息工程大学 | Neural network recommendation method and system based on meta-path fusion and heterogeneous network |
| CN115408621A (en) * | 2022-08-12 | 2022-11-29 | 中国测绘科学研究院 | Interest point recommendation method considering linear and nonlinear interaction of auxiliary information features |
| 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 |
| CN116842258A (en) * | 2023-07-03 | 2023-10-03 | 湖南大学 | Group recommendation method for feature enhancement based on meta-path context |
| CN116842260A (en) * | 2023-07-07 | 2023-10-03 | 天津大学 | Knowledge enhancement recommendation method based on graphic neural network multi-space interaction modeling |
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160157057A1 (en) * | 2014-04-22 | 2016-06-02 | Sk Planet Co., Ltd. | Device for recommending location in building by using fingerprint of access point, and method using same |
| US20210248461A1 (en) * | 2020-02-11 | 2021-08-12 | Nec Laboratories America, Inc. | Graph enhanced attention network for explainable poi recommendation |
| CN112069415A (en) * | 2020-08-13 | 2020-12-11 | 中国海洋大学 | Interest point recommendation method based on heterogeneous attribute network characterization learning |
| 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 |
| CN115408621A (en) * | 2022-08-12 | 2022-11-29 | 中国测绘科学研究院 | Interest point recommendation method considering linear and nonlinear interaction of auxiliary information features |
| CN115310005A (en) * | 2022-08-15 | 2022-11-08 | 中国人民解放军战略支援部队信息工程大学 | Neural network recommendation method and system based on meta-path fusion and heterogeneous network |
| CN116842258A (en) * | 2023-07-03 | 2023-10-03 | 湖南大学 | Group recommendation method for feature enhancement based on meta-path context |
| CN116842260A (en) * | 2023-07-07 | 2023-10-03 | 天津大学 | Knowledge enhancement recommendation method based on graphic neural network multi-space interaction modeling |
Non-Patent Citations (8)
| Title |
|---|
| WU, DAQING ET AL: "Composition-Enhanced Graph Collaborative Filtering for Multi-behavior Recommendation", 2021 21ST IEEE INTERNATIONAL CONFERENCE ON DATA MINING (ICDM 2021), 27 April 2022 (2022-04-27), pages 1427 - 1432 * |
| 任嘉睿等: "基于元图卷积的异质网络嵌入学习算法", 计算机研究与发展, vol. 59, no. 8, 31 August 2022 (2022-08-31), pages 1683 - 1693 * |
| 张云峰: "机器学习算法理论与应用", 30 June 2022, 中国海洋大学出版社, pages: 186 - 188 * |
| 方卿，徐丽芳: "出版学研究进展", 31 December 2017, 武汉大学出版社, pages: 415 * |
| 曹玖新;董羿;杨鹏伟;周涛;刘波;: "LBSN中基于元路径的兴趣点推荐", 计算机学报, no. 04, 30 April 2016 (2016-04-30), pages 675 - 684 * |
| 李全;许新华;刘兴红;林松;: "基于LBSN动态异构网络的时间感知兴趣点推荐", 计算机工程与应用, no. 11, 1 June 2020 (2020-06-01), pages 67 - 74 * |
| 王浩森: "基于异构信息网络表示学习的推荐模型研究", 中国优秀硕士学位论文全文数据库 (信息科技辑), 15 January 2023 (2023-01-15), pages 138 - 3436 * |
| 邢长征;赵宏宝;张全贵;郭亚兰;: "融合评论文本层级注意力和外积的推荐方法", 计算机科学与探索, vol. 14, no. 06, 30 June 2020 (2020-06-30), pages 947 - 957 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Qu et al. | Product-based neural networks for user response prediction over multi-field categorical data | |
| Wu et al. | Socialgcn: An efficient graph convolutional network based model for social recommendation | |
| Yuan et al. | Parameter-efficient transfer from sequential behaviors for user modeling and recommendation | |
| Yu et al. | Generating reliable friends via adversarial training to improve social recommendation | |
| CN112529168B (en) | GCN-based attribute multilayer network representation learning method | |
| Yang et al. | Autoencoder with invertible functions for dimension reduction and image reconstruction | |
| Zhuang et al. | Label information guided graph construction for semi-supervised learning | |
| Gou et al. | Multilevel attention-based sample correlations for knowledge distillation | |
| Deco et al. | An information-theoretic approach to neural computing | |
| Yang et al. | Skeletonnet: A hybrid network with a skeleton-embedding process for multi-view image representation learning | |
| Lin et al. | Attribute-Aware Convolutional Neural Networks for Facial Beauty Prediction. | |
| CN112883200A (en) | Link prediction method for knowledge graph completion | |
| CN113065974B (en) | Link prediction method based on dynamic network representation learning | |
| Fang et al. | Unbalanced incomplete multi-view clustering via the scheme of view evolution: Weak views are meat; strong views do eat | |
| Zhang et al. | Graph attention network with dynamic representation of relations for knowledge graph completion | |
| Qin et al. | High-quality temporal link prediction for weighted dynamic graphs via inductive embedding aggregation | |
| CN116992151A (en) | Online course recommendation method based on double-tower graph convolution neural network | |
| Zhang et al. | Selective knowledge transfer for cross-domain collaborative recommendation | |
| Xie et al. | Distributed semi-supervised learning algorithms for random vector functional-link networks with distributed data splitting across samples and features | |
| CN115098692A (en) | Cross-domain recommendation method and device, electronic equipment and storage medium | |
| US20240046127A1 (en) | Dynamic causal discovery in imitation learning | |
| CN117390300B (en) | Construction method and device of multi-channel interactive learning interest point recommendation model | |
| Yao et al. | Gckg: Novel gated convolutional embedding model for knowledge graphs | |
| Liu et al. | Deep latent-variable kernel learning | |
| CN117390300A (en) | Construction method and device of multi-channel interactive learning interest point recommendation model |
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 |