CN116186385A - Service recommendation method based on decoupling characterization learning and graph neural network - Google Patents
Service recommendation method based on decoupling characterization learning and graph neural network Download PDFInfo
- Publication number
- CN116186385A CN116186385A CN202211634971.2A CN202211634971A CN116186385A CN 116186385 A CN116186385 A CN 116186385A CN 202211634971 A CN202211634971 A CN 202211634971A CN 116186385 A CN116186385 A CN 116186385A
- Authority
- CN
- China
- Prior art keywords
- service
- user
- sequence
- interaction
- class
- 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 36
- 238000012512 characterization method Methods 0.000 title claims abstract description 17
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 14
- 230000003993 interaction Effects 0.000 claims abstract description 60
- 230000007774 longterm Effects 0.000 claims abstract description 19
- 239000011159 matrix material Substances 0.000 claims description 49
- 230000006870 function Effects 0.000 claims description 15
- 239000013598 vector Substances 0.000 claims description 12
- 230000002776 aggregation Effects 0.000 claims description 10
- 238000004220 aggregation Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 9
- 238000011176 pooling Methods 0.000 claims description 9
- 230000006399 behavior Effects 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 8
- 230000004927 fusion Effects 0.000 claims description 6
- 238000013527 convolutional neural network Methods 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000002457 bidirectional effect Effects 0.000 claims description 3
- 230000002452 interceptive effect Effects 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims 1
- 230000007547 defect Effects 0.000 abstract description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 101710181452 Retinal guanylyl cyclase 1 Proteins 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Physics & Mathematics (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 invention discloses a service recommendation method based on decoupling characterization learning and a graphic neural network, which takes a user interaction log recorded by online application as a data source to acquire a complete historical interaction sequence and a corresponding category sequence of a user, firstly, respectively capturing long-term stable interests and short-term dynamic interests of the user in a personalized degree, and then capturing popular interests of the user by utilizing service category information. The method aims at the problems that the implicit decoupling-based method relies on the number of potential interests set manually, is difficult to adapt to different application scenes, and does not consider different online application service class factors, long-term interests, short-term interests and popular interests of users are learned based on complete sequences, nearest sequences and class sequences respectively, the defect that the number of potential interests is set manually is avoided, the influence of different online application class factors is considered, the popular interests of the users are captured as the basis of the interpretation of recommendation results, and application providers are helped to adjust recommendation strategies in a targeted manner.
Description
Technical Field
The invention belongs to the technical field of recommendation systems, and particularly relates to a service recommendation method based on decoupling characterization learning and a graph neural network.
Background
In recent years, with the rapid popularization of internet technology, there is an increasing demand for online application services (such as online music, video streaming media, electronic shopping, etc.) in daily life. The recommendation system models the user portraits by using the historical behavior records of the users, helps online application providers to pertinently recommend services possibly interested for the users, and has great potential economic value. For example, the Internet cloud concert generates a user portrait according to the song listening record of the user, and further recommends potential songs with member identifications for the user to attract the user to upgrade members; the heaven shopping platform can recommend similar commodities according to the buying, searching and other actions of the user to help the user to purchase the commodities, so that the purchasing conversion rate is improved.
With the increasing deep learning technology, many current focus points for modeling user preferences based on user history behavior sequences are association relations between users and interactive services, for example, capturing long-term interests of users through sequence modes, capturing second-order and high-order links between users and services through building graph structures, and learning integral interest features of users. However, the behavior of the user is the result of a plurality of factors, and there is a long-term stable interest and a short-term random dynamic interest in individualization, a popular interest in the service popularity, and different interests have different degrees of influence on the future behavior of the user.
In order to capture the potential interests of a user in different aspects through the historical behavior record of the user, a plurality of methods adopt an implicit decoupling mode, firstly, the number of the potential interests is set as a premise of decoupling characterization learning, then the potential interests in different aspects are learned by utilizing characteristic learning methods such as a graph neural network, and finally the overall preference of the user is obtained. The existing method has the defects that the main distribution situation of potential interests is not analyzed on the whole, the number of the potential interests is difficult to adjust according to actual application recommended scenes, and the unique characteristics (such as category factors) of the online service are ignored, so that the economic benefits of online application providers are influenced.
Disclosure of Invention
Aiming at the problems that the method based on implicit decoupling characterization learning depends on the number of potential interests set manually, is difficult to adapt to different service scenes and does not consider different online application category factors, the invention provides a service recommendation method based on decoupling characterization learning and a graph neural network. Firstly, respectively capturing long-term stable interests and short-term dynamic interests of a user in a personalized degree, and then capturing popular interests of the user by using service class information.
The method comprises the following specific steps:
and (1) collecting a complete historical interaction sequence and a corresponding category sequence of a user by taking a user interaction log recorded by an online application as a data source. The complete history interaction sequence refers to interaction records (such as clicking, collecting, scoring and other actions) of services generated by a user, and the corresponding category sequence refers to a sequence constructed by category information corresponding to each service. Defining I as a set of all services, and the number of the services is n; defining C as a set of all service corresponding categories, and the number of the categories is k; define the historical interaction sequence of user u as t=t 1 ,t 2 ,…,t l-1 ,t l ,t l+1 Wherein t is i Records representing the ith interaction, including the service and corresponding class of tuples, i.e. t i ＝(I i ,C i ) L represents the sequence length, t l+1 Representing the next interaction to be predicted.
And (2) inputting the service set I and the class set C into an embedding layer to obtain embedded characteristic representation of each service and corresponding class. The embedded layer is characterized in that a multi-dimensional feature vector is obtained through a one-hot encoder and a multi-layer perceptron to represent each service or category, and embedded features obtained by the embedded layer are defined as two feature matrixes: x is X i ∈R n×d For serving feature matrix, X c ∈R k×d Is a category feature matrix, where n represents the number of services, k represents the number of categories, and d represents the feature dimension.
In order to introduce category factors to perform decoupling learning on user interests, a long-term interest modeling network and a short-term interest modeling network based on a user interaction sequence are designed, and a category-based user sub-interest modeling module comprises the following three steps:
and (3.1) firstly, constructing a hypergraph based on a service sequence in the complete history interaction sequence, and capturing long-term interests of a user after passing through a hypergraph convolution layer.
1) Defining a service hypergraph as GI E (V) i ,E i )，V i And E is i Respectively vertex set and superside set, and supergraph adjacent matrix H i ∈{0,1} m×n Creating supersides (supersides refer to special sides capable of connecting multiple vertexes) by using service sequences in historical behavior sequences of one user, and creating m supersides by m users, wherein the supergraph adjacency matrix H corresponds to services contained in the service sequences i The position will be set to 1, otherwise 0. E.g. the first user generates an interaction record with the first service, H i (0,0)＝1。
2) Vertex degree matrix D defining hypergraph iv ∈R n×n And a superside matrix D ie ∈R m×m They are diagonal matrices with respect to vertex degree and superside degree, respectively, where vertex V e V i Degree of (2)
Superedge E E i Degree of (2)
3) The definition hypergraph convolutional neural network is used for feature learning and information transmission, and the calculation mode between adjacent network layers is as follows
Wherein->
Representing service characteristics corresponding to a service sequence, W i ∈R m×m Representing a weight matrix, σ (·) representing a nonlinear activation function. And then using the average pooling strategy to obtain servicesThe characteristic is represented as follows:
wherein L represents the number of layers of the convolutional network, and finally, characteristic information in an average strategy aggregation sequence is used for obtaining long-term interest h of a user lo ：
Where l represents the user history interaction sequence length.
Step (3.2) the short-term interests of the user are then captured based on the service sequences in the recently historic interaction sequences, with the following specific calculation procedure.
1) T=t for the complete interaction sequence of the user 1 ,t 2 ,…,t l-1 ,t l Selecting the nearest interaction subsequence T s ＝t l-a ,…,t l-1 ,t l A represents the number of the selected interaction records, and can be set according to different application scenes. Then the service characteristic sequences corresponding to the subsequences pass through a bidirectional LSTM coding layer and are used for combining time information to capture short-term interest expression of the user
Wherein->
Representing an initial service feature representation. The feature matrix expressed by short-term interests of m users is H, and f represents a feature aggregation method.
2) Establishing a user graph GU by taking users as vertexes, wherein the condition of generating edge connection between vertexes is that the same nearest interaction service exists between the users, and capturing potential connection between different users by utilizing graph feature learning, wherein the specific propagation mode is as follows:
H (l+1) ＝σ(D -1 AH l )
wherein D and A are the degree matrices of the user graph GU, respectivelyAnd identity matrix, σ (·) is the activation function. Also using an average pooling strategy to obtain a user short-term interest representation h sh ∈H：
In order to adapt to the data characteristics of different applications, the method introduces service class information as a data base to capture the secondary interests generated by users aiming at class popularity. Similar to step (3.1), a class hypergraph GC E (V) is constructed from class sequences in the complete historical interaction sequence c ,E c )，V c And E is c Respectively vertex set and superside set, and supergraph adjacent matrix H c ∈{0,1} m×k . Defining class hypergraph convolution calculation:
wherein->
Representing class characteristics corresponding to class sequences, D cv ∈R k×k And D ce ∈R m×m Vertex degree matrix and superside degree matrix respectively representing class supergraph, W c ∈R m×m Representing a weight matrix, sigma (·) representing a nonlinear activation function, and obtaining the popular interest h of the user by averaging the pooling layers co 。
Step (4) based on the step (3), respectively making the learned user long-term interest h lo Short-term interest h sh And from the masses interest h co Feature fusion is carried out through the multi-interest aggregation layer, and a final user interest representation h is obtained, wherein the fusion mode is as follows:
h＝W[h lo ||h sh ||h co ]+b
wherein W and b represent the weight matrix and the bias term, respectively, [. Cn| - | - ] represents the concatenation of feature vectors in dimensions.
And (5) executing a service recommendation task based on the user interest h obtained in the step (4). For alternativesService set I, recommendation score z of each service is defined by service feature vector X ε X i And the user interest feature vector h, i.e. z i ＝x T h, normalizing z by softmax function to obtain final service score
The top-N services with the highest scores are then selected as candidate services in the recommendation list. The objective function of recommended task learning is expressed as cross entropy loss:
wherein the method comprises the steps of
Refers to the probability that the target service will interact given the interaction record sequence T.
Compared with the prior art, the invention has the following beneficial effects: the decoupling learning of the user interests by the method can learn the long-term interests, short-term interests and secondary interests of the user respectively, the defect of setting the number of potential interests manually is avoided to the greatest extent, the influence of different online application category factors is considered, the secondary interests of the user are captured as the basis of the interpretation of the recommendation results, and the application provider is helped to adjust the recommendation strategy in a targeted manner.
Drawings
Fig. 1 is a schematic diagram of a service recommendation method based on decoupling characterization learning and a neural network according to an embodiment of the present invention.
FIG. 2 is a block diagram of feature embedding based on a recent sequence in an embodiment of the present invention.
Detailed Description
It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other.
The technical scheme of the invention is further explained and illustrated below with reference to the attached drawings.
The embodiment provides a service recommendation method based on decoupling characterization learning and a graph neural network, and an overall execution framework of the method is shown in fig. 1.
And (1) collecting a complete historical interaction sequence and a corresponding category sequence of a user by taking a user interaction log recorded by an online application as a data source. The complete history interaction sequence refers to interaction records (such as clicking, collecting, scoring and other actions) of services generated by a user, and the corresponding category sequence refers to a sequence constructed by category information corresponding to each service. Defining I as a set of all services, and the number of the services is n; defining C as a set of all service corresponding categories, and the number of the categories is k; define the historical interaction sequence of user u as t=t 1 ,t 2 ,…,t l-1 ,t l ,t l+1 Wherein t is i Records representing the ith interaction, including the service and corresponding class of tuples, i.e. t i ＝(I i ,C i ) L represents the sequence length, t l+1 Representing the next interaction to be predicted.
And (2) inputting the service set I and the class set C into an embedding layer to obtain embedded characteristic representation of each service and corresponding class. The embedded layer is characterized in that a multi-dimensional feature vector is obtained through a one-hot encoder and a multi-layer perceptron to represent each service or category, and embedded features obtained by the embedded layer are defined as two feature matrixes: x is X i ∈R n×d For serving feature matrix, X c ∈R k×d Is a category feature matrix, where n and k represent the number of services and categories, respectively, and d represents the feature dimension.
In order to introduce category factors to perform decoupling learning on user interests, a long-term interest modeling network and a short-term interest modeling network based on a user interaction sequence are designed, and a category-based user sub-interest modeling module comprises the following three steps:
step (3.1) first based on the service sequence T in the complete historical interaction sequence i ＝I 1 ,I 2 ,…,I l-1 ,I l Constructing hypergraph, and capturing long-term interest of user after passing through hypergraph convolution layer to be expressed as h lo 。
1) Defining a service hypergraph as GI E (V) i ,E i )，V i And E is i Respectively vertex set and superside set, and supergraph adjacent matrix H i ∈{0,1} m×n The hyperedge is established by a service sequence in a historical behavior sequence of one user (the hyperedge refers to a special edge capable of connecting a plurality of vertexes), m hyperedges are established by m users, the Hi position of a hypergraph adjacent matrix corresponding to the service contained in the service sequence is set to be 1, and otherwise, the Hi position is set to be 0. If the first user generates an interaction record with the first service, H i (0, 0) =1. For example two users u 1 And u 2 The service sequences in the history interaction records of (1) are { I }, respectively 1 ,I 2 ,I 2 ,I 3 Sum { I } 2 ,I 4 ,I 2 ,I 5 The corresponding service hypergraph adjacency matrix is:
2) Vertex degree matrix D defining hypergraph iv ∈R n×n And a superside matrix D ie ∈R m×m They are diagonal matrices with respect to vertex degree and superside degree, respectively, where vertex V e V i Degree of (2)
Superedge E E i Degree of (2)
3) The definition hypergraph convolutional neural network is used for feature learning and information transmission, and the calculation mode between adjacent network layers is as follows
Wherein->
Representing service characteristics corresponding to a service sequence, W i ∈R m×m Representation rightsThe heavy matrix, σ (·) represents the nonlinear activation function. The average pooling strategy is then used to derive a feature representation of the service:
where L represents the number of convolutionally networked layers,
representing a service feature matrix of a first layer, and finally obtaining long-term interest h of a user by using feature information in an average strategy aggregation sequence lo ：
Where l represents the length of the user history interaction sequence, X i,t A feature representation representing a t-th service in the sequence of services.
Step (3.2) the short-term interests of the user are then captured based on the service sequences in the recent historical interaction sequence, the session embedding aggregation module being as shown in fig. 2, the specific calculation process being as follows.
1) T=t for the complete interaction sequence of the user 1 ,t 2 ,…,t l-1 ,t l Selecting the nearest interaction subsequence T s ＝t l-a ,…,t l-1 ,t l A represents the number of the selected interaction records, and can be set according to different application scenes. Then the service characteristic sequences corresponding to the subsequences pass through a bidirectional LSTM coding layer and are used for combining time information to capture short-term interest expression of the user
Wherein->
Representing an initial service feature representation. The feature matrix of short-term interest representation of m users is H E R m×d F represents a bi-directional LSTM feature aggregation method.
2) Establishing user graph GU with user as vertex, and generating edge connection between vertices is based on the condition that the same nearest interaction service exists between users, such as user u 1 And u 2 Is { I }, respectively 2 ,I 3 Sum { I } 2 ,I 5 Presence of the same service I 2 Thus user u 1 And u 2 Edge connections may be created. The potential links between different users are captured by graph characterization learning, and the specific propagation modes are as follows:
H (l+1) ＝σ(D -1 AH l )
wherein D is E R m×m And A.epsilon.R m×m The degree matrix and identity matrix of the user graph GU, respectively, and σ (·) is the activation function. Also using an average pooling strategy to obtain a user short-term interest representation h sh ∈H：
Where L represents the number of layers of the user graph convolutional neural network,
representing a user short-term interest feature representation of the first layer.
In order to adapt to the data characteristics of different applications, the method introduces service class information as a data base to capture the secondary interest h generated by the user aiming at the class popularity co . Similar to step (3.1), to complete the category sequence T in the historical interaction sequence c ＝C 1 ,C 2 ,…,C l-1 ,C l Build class hypergraph GC epsilon (V) c ,E c )，V c And E is c Respectively vertex set and superside set, and supergraph adjacent matrix H c ∈{0,1} m×k . For example two users u 1 And u 2 The category sequences in the history interaction record of (C) are respectively { C } 1 ,C 1 ,C 1 ,C 2 Sum { C } 2 ,C 1 ,C 2 ,C 3 The corresponding class hypergraph adjacency matrix is:
defining class hypergraph convolution calculation:
wherein->
Representing class characteristics corresponding to class sequences, D cv ∈R k×k And D ce ∈R m×m Vertex degree matrix and superside degree matrix respectively representing class supergraph, W c ∈R m×m Representing a weight matrix, sigma (·) representing a nonlinear activation function, and obtaining the popular interest h of the user by averaging the pooling layers co 。
Step (4) based on the step (3), respectively making the learned user long-term interest h lo Short-term interest h sh And from the masses interest h co Feature fusion is carried out through the multi-interest aggregation layer, and a final user interest representation h is obtained, wherein the fusion mode is as follows:
h＝W[h lo ||h sh ||h co ]+b
wherein W is E R d×3d And b.epsilon.R d×1 Respectively representing a weight matrix and a bias term, [ carryingout I carrying out]Representing a concatenation of feature vectors in dimensions.
And (5) executing a service recommendation task based on the user interest h obtained in the step (4). For alternative service set I, recommendation score z for each service is determined by service feature vector xε X i And the user interest feature vector h, i.e. z i ＝x T h, normalizing z by softmax function to obtain final service score
The top-N services with the highest scores are then selected as candidate services in the recommendation list. The objective function of recommended task learning is expressed as crossoverEntropy loss:
wherein the method comprises the steps of
Refers to the probability that the target service will interact given the interaction record sequence T. In terms of the interpretability of the results, e.g. user u 1 The service sequence and class sequence of (1) are { I }, respectively 1 ,I 2 ,I 3 Sum { C } 1 ,C 1 ,C 2 If user u 1 The article actually next to interact is I 4 (category C 1 ) The results cannot be interpreted based on the service history alone, but user u can be seen from the category data 1 Frequent interaction category C 1 And the user can be focused on interpreting the recommendation from the public interests.
According to the method, long-term interests and short-term interests of a user are respectively captured on the basis of user history interaction record data and on the basis of decoupling characterization learning, then service class popularity factors are introduced, class hypergraph convolutional neural networks are constructed to capture popular interests of the user for service classes, and the capability of a recommendation model for adapting to different application scenes and interpretability of results is improved.
Claims (6)
1. The service recommendation method based on decoupling characterization learning and graph neural network is characterized by comprising the following steps:
s1, collecting a complete historical interaction sequence of a user and a corresponding class sequence, wherein the complete historical interaction sequence refers to an interaction record of the user for services, which is generated by the user, and the corresponding class sequence refers to a class information construction sequence corresponding to each service;
s2, inputting the set of the services and the set of the categories into an embedding layer to obtain each service and the embedded characteristic representation of the corresponding category, wherein the embedding layer is obtained through a one-hot encoder and a multi-layer perceptronThe multidimensional feature vector represents each service or category, and the embedded features obtained by defining the embedded layer are expressed as two feature matrices: x is X i ∈R n×d For serving feature matrix, X c ∈R k×d The method is a category feature matrix, wherein n represents the number of services, k represents the number of categories, and d represents feature dimensions;
s3, constructing a long-term interest and short-term interest modeling network based on the user interaction sequence, and modeling a class-based user secondary interest modeling module
S3-1, building a hypergraph based on a service sequence in a complete history interaction sequence, and capturing long-term interests of a user after passing through a hypergraph convolution layer;
s3-2, capturing short-term interests of the user based on service sequences in the latest historical interaction sequences;
s3-3, introducing service class information as a data base to capture the secondary interests generated by the users aiming at the class popularity;
s4, based on the step S3, respectively carrying out long-term interest h on the learned users lo Short-term interest h sh And from the masses interest h co Feature fusion is carried out through the multi-interest aggregation layer, and a final user interest representation h is obtained, wherein the fusion mode is as follows:
h＝W[h lo ||h sh ||h co ]+b
where W and b represent the weight matrix and the bias term respectively, [ carryingout I carrying out representative characteristics of ] splicing vectors in dimensions;
s5, based on the user interest h learned in the step S4, executing a service recommendation task, and aiming at an alternative service set I, recommending scores z of each service are formed by service feature vectors X epsilon X i And the user interest feature vector h, i.e. z i ＝x T h, normalizing z by softmax function to obtain final service score
Then, top-N services with highest scores are selected as candidate services in a recommendation list, and an objective function of recommendation task learning is expressed as cross entropy loss:
wherein the method comprises the steps of
Refers to the probability that the target service will interact given the interaction record sequence T.
2. The service recommendation method based on decoupling characterization learning and graph neural network according to claim 1, wherein the complete history interaction sequence and the corresponding category sequence of the user use a user interaction log recorded by an online application as a data source, and the interaction record comprises clicking, collecting and scoring.
3. The service recommendation method based on decoupling characterization learning and graph neural network according to claim 1, wherein I is defined as a set of all services, and the number of services is n; defining C as a set of all service corresponding categories, and the number of the categories is k; define the historical interaction sequence of user u as t=t 1 ，t 2 ，...，t l-1 ，t l ，t l+1 Wherein t is i Records representing the ith interaction, including the service and corresponding class of tuples, i.e. t i ＝(I i ，C i ) L represents the sequence length, t l+ Representing the next interaction to be predicted.
4. The service recommendation method based on decoupling characterization learning and graph neural network according to claim 3, wherein the step S3-1 is specifically as follows:
s3-1-1, defining service hypergraph as GI E (V) i ，E i )，V i And E is i Respectively vertex set and superside set, and supergraph adjacent matrix H i ∈{0，1} m×n Creating superedges with service sequences in a user's historical behavior sequence, said superedgesThe edges are special edges capable of connecting a plurality of vertexes, m users can establish m hyperedges, the position of a hypergraph adjacent matrix H corresponding to the service contained in the service sequence is set to be 1, and otherwise, the position is set to be 0;
s3-1-2, defining vertex degree matrix D of hypergraph iv ∈R n×n And a superside matrix D ie ∈R m×m Vertex degree matrix D iv ∈R n×n And a superside matrix D ie ∈R m×m Diagonal matrices for vertex and superedge, respectively, where vertex V e V i Degree of (2)
Superedge E E i Degree of (1)>
S3-1-3, defining a hypergraph convolutional neural network for feature learning and information propagation, wherein the calculation mode between adjacent network layers is as follows
Wherein->
Representing service characteristics corresponding to a service sequence, W i Representing a weight matrix, σ (·) representing a nonlinear activation function, and then using an average pooling strategy to obtain a feature representation of the service:
wherein L represents the number of layers of the convolutional network, and finally, characteristic information in an average strategy aggregation sequence is used for obtaining long-term interest h of a user lo ：
Where l represents the user history interaction sequence length.
5. The service recommendation method based on decoupling characterization learning and graph neural network according to claim 4, wherein the specific method of step S3-2 is as follows:
s3-2-1, t=t for the complete interaction sequence of the user 1 ，t 2 ，...，t l-1 ，t l Selecting the nearest interaction subsequence T s ＝t l-a ，...，t l-1 ，t l A represents the number of selected interactive records, and then the service feature sequences corresponding to the subsequences pass through a bidirectional LSTM coding layer for capturing short-term interest representation of the user in combination with time information
Wherein->
Representing initial service feature representation, wherein a feature matrix of short-term interest representation of m users is H, and a feature aggregation method of f tables;
s3-2-2, establishing a user graph GU by taking users as vertexes, wherein the condition of generating edge connection between the vertexes is that the same nearest interaction service exists between the users, and capturing potential relations between different users by utilizing graph feature learning, wherein the specific propagation mode is as follows:
H (l+1) ＝σ(D -1 AH l )
wherein D and A are the degree matrix and identity matrix of the user graph GU, respectively, and sigma (&) is the activation function, and the average pooling strategy is also used to obtain the short-term interest representation h of the user sh ∈H：
6. The service recommendation method based on decoupling characterization learning and graph neural network according to claim 5, wherein the specific method of step S3-3 is as follows:
building a class hypergraph GC epsilon (V) from class sequences in a complete historical interaction sequence c ，E c )，V c And E is c Respectively vertex set and superside set, and supergraph adjacent matrix H c ∈{0，1} m×k Defining class hypergraph convolution calculation:
wherein->
Representing class characteristics corresponding to class sequences, D cv ∈R k×k And D ce ∈R m×m Vertex degree matrix and superside degree matrix respectively representing class supergraph, W c Representing a weight matrix, sigma (·) representing a nonlinear activation function, and obtaining the popular interest h of the user by averaging the pooling layers co 。/>
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211634971.2A CN116186385A (en) | 2022-12-19 | 2022-12-19 | Service recommendation method based on decoupling characterization learning and graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211634971.2A CN116186385A (en) | 2022-12-19 | 2022-12-19 | Service recommendation method based on decoupling characterization learning and graph neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116186385A true CN116186385A (en) | 2023-05-30 |
Family
ID=86437430
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211634971.2A Pending CN116186385A (en) | 2022-12-19 | 2022-12-19 | Service recommendation method based on decoupling characterization learning and graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116186385A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117131438A (en) * | 2023-10-27 | 2023-11-28 | 深圳市迪博企业风险管理技术有限公司 | Litigation document analysis method, model training method, device, equipment and medium |
-
2022
- 2022-12-19 CN CN202211634971.2A patent/CN116186385A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117131438A (en) * | 2023-10-27 | 2023-11-28 | 深圳市迪博企业风险管理技术有限公司 | Litigation document analysis method, model training method, device, equipment and medium |
| CN117131438B (en) * | 2023-10-27 | 2024-02-13 | 深圳市迪博企业风险管理技术有限公司 | Litigation document analysis method, model training method, device, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110162703B (en) | Content recommendation method, training device, content recommendation equipment and storage medium | |
| Al-Shamri et al. | Fuzzy-genetic approach to recommender systems based on a novel hybrid user model | |
| CN110362738B (en) | Deep learning-based individual recommendation method combining trust and influence | |
| CN111061962B (en) | Recommendation method based on user scoring analysis | |
| CN110795619A (en) | Multi-target-fused educational resource personalized recommendation system and method | |
| WO2020135642A1 (en) | Model training method and apparatus employing generative adversarial network | |
| CN113590965B (en) | Video recommendation method integrating knowledge graph and emotion analysis | |
| CN113361928B (en) | Crowd-sourced task recommendation method based on heterogram attention network | |
| CN107391577B (en) | Work label recommendation method and system based on expression vector | |
| CN115408621B (en) | Interest point recommendation method considering auxiliary information characteristic linear and nonlinear interaction | |
| CN111967946B (en) | Commodity recommendation method and system based on user-oriented multi-relation information network | |
| CN115270004B (en) | Educational resource recommendation method based on field factor decomposition | |
| CN116186385A (en) | Service recommendation method based on decoupling characterization learning and graph neural network | |
| CN111324773A (en) | Background music construction method and device, electronic equipment and storage medium | |
| CN115618024A (en) | Multimedia recommendation method and device and electronic equipment | |
| Wang et al. | Online course recommendation algorithm based on multilevel fusion of user features and item features | |
| CN113095883A (en) | Video payment user prediction method and system based on deep cross attention network | |
| CN117495511A (en) | Commodity recommendation system and method based on contrast learning and community perception | |
| CN113962417A (en) | Video processing method and device, electronic equipment and storage medium | |
| CN113742561A (en) | Video recommendation method and device, electronic equipment and storage medium | |
| CN116541592A (en) | Vector generation method, information recommendation method, device, equipment and medium | |
| CN115310004A (en) | Graph nerve collaborative filtering recommendation method fusing project time sequence relation | |
| CN115545834A (en) | Personalized service recommendation method based on graph neural network and metadata | |
| CN113742562B (en) | Video recommendation method and device, electronic equipment and storage medium | |
| CN115689639A (en) | Commercial advertisement click rate prediction method based on deep 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 |