CN116701611A - Recommendation method and system for learning knowledge graph fusing interaction attention - Google Patents
Recommendation method and system for learning knowledge graph fusing interaction attention Download PDFInfo
- Publication number
- CN116701611A CN116701611A CN202310602037.0A CN202310602037A CN116701611A CN 116701611 A CN116701611 A CN 116701611A CN 202310602037 A CN202310602037 A CN 202310602037A CN 116701611 A CN116701611 A CN 116701611A
- Authority
- CN
- China
- Prior art keywords
- knowledge graph
- vector
- module
- recommendation
- user
- 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
- 230000003993 interaction Effects 0.000 title claims abstract description 65
- 238000000034 method Methods 0.000 title claims abstract description 33
- 230000006870 function Effects 0.000 claims abstract description 52
- 238000012549 training Methods 0.000 claims abstract description 45
- 239000011159 matrix material Substances 0.000 claims abstract description 38
- 238000013507 mapping Methods 0.000 claims abstract description 7
- 239000013598 vector Substances 0.000 claims description 107
- 230000006835 compression Effects 0.000 claims description 29
- 238000007906 compression Methods 0.000 claims description 29
- 230000002452 interceptive effect Effects 0.000 claims description 9
- 238000013528 artificial neural network Methods 0.000 claims description 8
- 238000002360 preparation method Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 238000000605 extraction Methods 0.000 claims description 4
- 230000017105 transposition Effects 0.000 claims description 4
- 230000000694 effects Effects 0.000 abstract description 5
- 238000012986 modification Methods 0.000 description 4
- 230000004048 modification Effects 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 101150071716 PCSK1 gene Proteins 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000001914 filtration 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
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/335—Filtering based on additional data, e.g. user or group profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- 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
- G06N3/042—Knowledge-based neural networks; Logical representations of neural 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
- G06N3/04—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
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Animal Behavior & Ethology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a recommendation method and a recommendation system for learning knowledge maps by fusing interaction attention, wherein the recommendation method comprises the following steps: acquiring interaction information of a user-object and information of a knowledge graph, constructing an interaction matrix of the user-object according to a history record of scoring the object by the user, and constructing the knowledge graph after mapping is established according to an entity of a knowledge graph data set and a recommendation data set; inputting the obtained object features and the head entity features of the knowledge graph into a multi-task learning sharing module, training the knowledge graph module in an alternate training mode, and obtaining a scoring function of the knowledge graph triplet through a similarity function; and training a recommendation module by using an alternate training mode based on the multi-task learning sharing module, and finally obtaining the final prediction probability of the user participating in the project by using a prediction function. The invention can fully excavate the information contained in the knowledge graph module and the recommendation module, effectively relieve the problem that the upper text information, the lower text information and the data sparsity cannot be utilized, and further improve the actual recommendation effect.
Description
Technical Field
The invention belongs to the technical field of artificial intelligence, and particularly relates to a recommendation method and a recommendation system for learning knowledge maps by fusing interaction attention.
Background
Before the advent of big data and artificial intelligence, traditional recommendation models are generally based on content recommendation, collaborative filtering, model-based recommendation and the like, but with the increasing of data volume explosion and the increasing of data types, the problems of data sparseness and cold start exposed by traditional recommendation methods are more and more obvious, so that the traditional recommendation methods are not attractive. With the rapid development of deep learning, a deep learning algorithm is successfully integrated in the recommendation model, and the recommendation model integrated with the deep learning has stronger learning ability and expression ability, thereby attracting attention at home and abroad.
Disclosure of Invention
Aiming at the technical problems in the prior art, the invention provides a recommendation method and a recommendation system for learning knowledge maps by fusing interaction attention. Compared with the traditional knowledge graph enhancement recommendation method, the method introduces a multi-task learning sharing module, interactively trains the knowledge graph and the recommendation, and relieves the problems of data sparseness and cold start in the traditional recommendation method.
According to a first aspect of the present invention, there is provided a recommendation method for learning knowledge graph with integrated interaction attention, including a multi-task learning sharing module, including the steps of:
step 1: acquiring interaction information of a user-object from a recommendation data set, and acquiring knowledge-graph information from a knowledge-graph data set; constructing an interaction matrix of the user-object according to the historical record of the user scoring object, and constructing a knowledge graph after mapping is established according to the knowledge graph data set and the entity of the recommendation data set; the user-article interaction information comprises a set of users and articles and a historical record of scoring articles by the users, and the knowledge graph consists of a head entity-relation-tail entity triplet;
step 2: inputting article characteristics and head entity characteristics obtained from the user-article interaction information into a multi-task learning sharing module;
step 3: training a knowledge graph module by an alternate training mode based on the multi-task learning sharing module, and obtaining a scoring function of the knowledge graph triplet by a similarity function;
step 4: and training a recommendation module by using an alternate training mode based on the multi-task learning sharing module, and finally obtaining the final prediction probability of the user participating in the project by using a prediction function.
On the basis of the technical scheme, the invention can also make the following improvements.
Optionally, the multi-task learning sharing module is composed of a plurality of feature attention cross compression modules, wherein the feature attention cross compression modules comprise feature interaction attention units and cross compression units, and the object features and the head entity features are learned by the feature interaction attention units to obtain object features and entity features which are integrated with the context information, so that preparation is made for interactive training of the recommendation module and the knowledge graph module.
Optionally, the recommendation module includes a plurality of feature extraction layers and a prediction layer connected in sequence.
Optionally, the inputting the item feature and the head entity feature into the multi-task learning sharing module includes:
step 2.1: for each input vector of candidate articles and head entities, multiplying the vector by a trainable parameter matrix to obtain a query vector and an address vector of each vector;
step 2.2: calculating similarity scores of the query vector and the address vector, and then applying a normalized exponential function to the similarity scores to obtain respective weight coefficient scores;
step 2.3: weighting and summing the weight coefficient scores and the article vectors or the head entity vectors respectively to obtain article self-attention scores and head entity self-attention scores respectively;
step 2.4: performing point multiplication on the self-attention score of the article and the self-attention score of the head entity processed by the activation function respectively;
step 2.5: adding the point multiplication value and the initial feature vector to obtain an article vector and a head entity vector respectively;
step 2.6: the article vector and the head entity vector are subjected to characteristic crossing in a crossing compression unit to obtain a crossing multiplication matrix;
step 2.7: transposition is carried out on the cross multiplication matrix to obtain a transposition cross multiplication matrix;
step 2.8: and compressing the crossed feature matrix and the transposed matrix by calculation, and calculating and outputting the project and the entity feature vector of the next layer.
Optionally, in step 2.2, a normalized exponential function is applied to the similarity score, resulting in respective weight coefficient scores expressed as:
wherein score1 and score2 are the weight coefficient score of the item and the weight coefficient score of the entity, respectively; d, d k Is the query vector Q 1 、Q 2 And address vector K 1 T 、K 2 T Vector dimensions of the matrix.
Optionally, in step 2.6, the article feature vector and the head entity feature vector are subjected to feature intersection in an intersection compression unit to obtain an intersection multiplication matrix, and the intersection multiplication matrix in d×d dimensions is constructed by calculating an outer product of the article vector and the head entity vector, and is expressed as:
wherein ,BL ∈R dxd Is the L-layer feature cross multiplication matrix and d is the dimension of the embedded vector.Is B L The interaction of the item v with each feature of its associated header entity h.
Optionally, in step 2.8, the item and entity feature vector of the next layer are calculated and output as:
wherein , and />Is a trainable compression unit weight vector and bias term representing the L-th layer for the item vector v and the head entity vector h.
Optionally, in step 3, the obtaining the score function of the knowledge-graph triplet through the similarity function includes:
step 3.1: for a given triplet, the head entity obtains the high-order characteristic of the head entity through a multi-task learning sharing module;
step 3.2: extracting implicit cross characteristics of a relational entity by using a multi-layer neural network;
step 3.3: the head entity characteristics and the relation characteristics are embedded into a function through a knowledge graph to predict a prediction vector with a tail entity as a tail;
step 3.4: and obtaining a score function of the triplet through the similarity function.
Optionally, in step 4, the training the recommendation module by the multi-task learning sharing module through an alternate training manner, and finally obtaining the final prediction probability of the user participation item through the prediction function includes:
step 4.1: extracting compression characteristics of hidden layers of a user through a multi-layer neural network layer of an L layer after characteristic embedding for initial characteristic vectors of the user;
step 4.2: for the original input vector of the article, calculating to obtain the potential characteristics of the article through a multi-task learning sharing module;
step 4.3: and calculating a final prediction probability of the user participation project through the prediction function.
According to a second aspect of the present invention, there is provided a recommendation system for learning knowledge graph fusing interactive attention, comprising:
the data acquisition module is used for acquiring interaction information of the user and the object from the recommendation data set and acquiring knowledge-graph information from the knowledge-graph data set; constructing an interaction matrix of the user-object according to the historical record of the user scoring object, and constructing a knowledge graph after mapping is established according to the knowledge graph data set and the entity of the recommendation data set; the user-article interaction information comprises a set of users and articles and a historical record of scoring articles by the users, and the knowledge graph consists of a head entity-relation-tail entity triplet;
the interactive training preparation module is used for inputting the article characteristics and the head entity characteristics obtained from the interactive information of the user-article into the multi-task learning sharing module;
the knowledge graph training module is used for training the knowledge graph module in an alternating training mode based on the multi-task learning sharing module and obtaining a score function of the triplet through the similarity function;
and the recommendation module training module is used for training the recommendation module based on the multi-task learning sharing module in an alternating training mode, and finally obtaining the final prediction probability of the user participation project through the prediction function.
The invention has the technical effects and advantages that:
according to the invention, the knowledge graph is combined with the recommendation method, so that the problem of inaccurate recommendation results caused by the problem of sparsity of interaction data in the traditional recommendation method is solved, a user can be quickly helped to filter useful information from huge information, and the recommendation accuracy is effectively improved.
The invention carries out joint training of the knowledge graph and the recommendation module by the multi-task learning sharing module. The problem of data sparsity of the traditional recommendation method is solved, and meanwhile, the problem of cold start is effectively avoided; the problems of unavailable use of the context information and insufficient interaction can be effectively solved, so that the actual recommendation effect is improved.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.
Drawings
Fig. 1 is a model architecture diagram of a recommendation method of learning knowledge graph with integrated interaction attention according to an embodiment of the present invention;
fig. 2 is a model diagram of a feature attention cross compression unit of a recommendation method of learning knowledge graph with integrated interaction attention according to an embodiment of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It can be appreciated that, based on the defects in the background technology, the embodiment of the invention provides a recommendation method for learning knowledge maps by fusing interaction attention, which comprises the following steps:
step 1: information for acquiring information of interaction of a user-item and knowledge graph data from a recommendation data set to acquire knowledge graph, wherein the information of interaction of the user-item comprises a set of the user and the item and a history record of scoring the item by the user, U= { U1, U2, & gt, um } and V= { V1, V2, & gt, vn } respectively represent the set of the user and the item, and a user-item interaction matrix Y= { Y is constructed according to the history record of scoring the item by the user uv U e U, V e V, where y uv A value of 1 indicates that there is an interaction between user u and item v, such as clicking, viewing, browsing, etc. Wherein y is uv The value of 0 indicates that there is no interaction between the user u and the item v, and a knowledge graph G is constructed after mapping is established according to the entities of the knowledge graph dataset and the recommendation dataset, the knowledge graph G being composed of a large number of entity-relationship-entity triples (h, r, t). Where h E, R E R and t E represent the head, relationship and tail entities of the knowledge graph, respectively, where E and R represent the set of entities and relationships in the knowledge graph G.
Step 2: and inputting the object characteristics and the head entity characteristics obtained from the user-object interaction information into the multi-task learning sharing module.
As shown in fig. 2, the multi-task learning sharing module is composed of L feature attention cross compression modules, wherein the feature attention cross compression modules comprise feature interaction attention units and cross compression units, the feature interaction attention units can learn the features of the object and the features of the head entity, distinguish the importance of different cross features, learn the context semantic information of the head entity, and enhance the interpretability of the model. Connectivity of higher order cross features between item features and header entity features may be captured at the cross compression unit.
It should be noted that the L feature attention cross compression modules may mean one or more, i.e., one or more. In step 2, specifically, the method includes:
step 2.1: for each candidate item entered and vector of head entity features, denoted v and h, respectively using a trainable parameter matrix and />Multiplied by it, which is expressed as:
wherein ,Q1 、K 1 、V 1 Is a trainable matrix of the article vector, representing a query vector, an address vector and a value vector matrix corresponding to the article vector; q (Q) 2 、K 2 、V 2 Is a trainable matrix of head entity vectors, representing a matrix of query vectors, address vectors, and value vectors corresponding to the head entity vectors.
Step 2.2: using the query vector and the address vector of step 2.1, calculating the relevance of the query vector and the address vector, namely the similarity score of the two vectors, and then applying a normalized exponential function to the similarity score to obtain respective weight coefficient scores, which are expressed as:
wherein ,dk Is the query vector Q 1 、Q 2 And address vector K 1 T 、K 2 T Vector dimensions of the matrix; score1 and score2 are the weight coefficient score of the item and the weight coefficient score of the entity, respectively.
Step 2.3: and (2) respectively carrying out weighted summation on the weight coefficient scores obtained in the step (2.2) and the article vector (h) or the head entity vector (v) to obtain a final article self-attention score (Atten 1) and a head entity self-attention score (Atten 2), wherein the final article self-attention score (Atten 2) is expressed as follows:
Atten1＝score1·h
Atten2＝score2·v
step 2.4: and (3) performing point multiplication on the self-attention score obtained in the step (2.3) and the self-attention score processed by the activation function (sigmoid) to obtain an article vector v1 and a head entity vector h1 respectively. The normalized exponential function here corresponds to a gating effect that can be used to select more important features from the features, expressed as:
h1＝Atten1·sigmoid(Atten1)
v1＝Atten2·sigmoid(Atten2)
step 2.5: and (3) adding the article vector and the head entity vector obtained in the step (2.4) with the initial feature vector respectively to finally obtain an article vector v and a head entity vector h. It is expressed as:
v＝add(Att1,v)
h＝add(Att2,h)
step 2.6: and (3) carrying out feature cross compression on the head entity vector and the article vector obtained in the step (2.5) in a cross compression layer by a cross compression unit, and setting the article vector of L layers by the article feature vector v and the head entity feature vector h obtained by the feature cross attention unit in the cross compression layerHead entity vector->Then the layer L cross-multiplication matrix B L The calculation of (2) is expressed as:
wherein BL ∈R dxd Is the layer L feature interaction multiplication matrix and d is the dimension of the embedded vector.Is B L The interaction of the item v with each feature of its associated header entity h.
Step 2.7: multiplying the interactions in step 2.6 by matrix B L Transposed to obtain transposed cross-multiplication matrixExpressed as:
step 2.8: will multiply matrix B L And transposed cross-multiplication matrixRecovering the dimension of the multiplication matrix from d x d to d x 1 by the weight vector to obtain the article vector v of the L+1 layer L+1 Sum-head entity vector h L+1 It is expressed as:
wherein , and />Is a trainable compression unit weight vector and bias term representing the L-th layer for the item vector v and the head entity vector h.
Step 3: the knowledge graph module is trained by the alternative training mode based on the multi-task learning sharing module.
In this embodiment, the specific implementation includes the following sub-steps:
step 3.1: the implicit cross features of the relational entity r are extracted by using a multi-layer neural network, and are expressed as follows:
rL＝M(M...M(r))＝M L (r)
wherein rL represents a feature of a relationship entity of layer L, M L () Representing a full connection layer function.
Step 3.2: extracting implicit cross characteristics of a relational entity by using a multi-layer neural network;
it is expressed as:
rL＝M(M...M(r))＝M L (r)
step 3.3: first, the head entity characteristics and the relation characteristics of the L layerSplicing, and obtaining a tail entity t prediction vector through a multi-layer neural networkIt is expressed as:
step 3.4: then a score function of the triplet is obtained by a similarity function, wherein the similarity function is t andafter inner product is made, a sigmoid function is then passed, which is expressed as:
where score (h, r, t) is the similarity score and σ is the activation function.
Step 4: and training a recommendation module by using an alternate training mode based on the multi-task learning sharing module, and finally obtaining the final prediction probability of the user participating in the project by using a prediction function.
The recommendation module includes L feature extraction layers and one prediction layer connected in sequence. The L feature extraction layers may be represented as a plurality.
In this embodiment, the training recommendation module based on the multi-task learning sharing module through an alternate training mode, and finally obtaining the final prediction probability of the user participation item through the prediction function specifically includes the following sub-steps:
step 4.1: for an initial feature vector of a user u, after feature embedding, extracting a compression feature uL of a hidden layer of the user through a neural network layer of an L layer, wherein the compression feature uL is expressed as follows:
uL＝M(M(...M(u)))＝M L (u)
step 4.2: for the original input vector v of the item v, learning by multitaskingSharing module for calculating potential characteristics of the last output article of the L layerIt is expressed as:
wherein , and />Is a trainable compression unit weight vector and bias term representing the time of the layer L pressing operation against the item vector v.
Step 4.3: and calculating to obtain the final prediction probability of the user u participating in the item v through the prediction function.
It is expressed as:
it should be noted that the prediction function here is to obtain the prediction score through a sigmoid function after the inner product of the user vector and the item vector.
In summary, compared with the traditional recommendation method with the enhanced knowledge graph, the method introduces a multi-task learning sharing module, interactively trains the knowledge graph and the recommendation, and relieves the problems of data sparseness and cold start in the traditional recommendation method. In the invention, firstly, in a multi-task learning module, the object features and the head entity features are subjected to feature interaction through a feature interaction attention unit, so that finer granularity features are learned, and after the object features and the head entity features which pass through the interaction attention module pass through a cross compression unit, the knowledge graph and the recommendation module are trained alternately in sequence. The invention can fully excavate the information contained in the knowledge graph module and the recommendation module, effectively relieve the problem that the upper text information, the lower text information and the data sparsity cannot be utilized, and further improve the actual recommendation effect.
According to a second aspect of the present invention, an embodiment of the present invention further provides a recommendation system for learning knowledge maps with integrated interaction attention, including:
the data acquisition module is used for acquiring interaction information of the user and the object from the recommendation data set and acquiring knowledge-graph information from the knowledge-graph data set; constructing an interaction matrix of the user-object according to the historical record of the user scoring object, and constructing a knowledge graph after mapping is established according to the knowledge graph data set and the entity of the recommendation data set; the user-article interaction information comprises a set of users and articles and a historical record of scoring articles by the users, and the knowledge graph consists of a head entity-relation-tail entity triplet;
the interactive training preparation module is used for inputting article characteristics and head entity characteristics of the knowledge graph obtained in the interactive information of the user-article into the multi-task learning sharing module;
the knowledge graph training module is used for training the knowledge graph module in an alternating training mode based on the multi-task learning sharing module and obtaining a score function of the triplet through the similarity function;
and the recommendation module training module is used for training the recommendation module based on the multi-task learning sharing module in an alternating training mode, and finally obtaining the final prediction probability of the user participation project through the prediction function.
It may be understood that the recommendation system for learning knowledge graph with integrated interaction attention provided by the present invention corresponds to the recommendation method for learning knowledge graph with integrated interaction attention provided in the foregoing embodiments, and the relevant technical features of the recommendation system for learning knowledge graph with integrated interaction attention may refer to the relevant technical features of the recommendation method for learning knowledge graph with integrated interaction attention, which are not described herein again.
It should be understood that parts of the specification not specifically set forth herein are all prior art.
It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.
Finally, it should be noted that: the foregoing description is only illustrative of the preferred embodiments of the present invention, and although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments described, or equivalents may be substituted for elements thereof, and any modifications, equivalents, improvements or changes may be made without departing from the spirit and principles of the present invention.
Claims (10)
1. A recommendation method of learning knowledge graph integrating interaction attention, which comprises a multi-task learning sharing module and is characterized in that,
step 1: acquiring interaction information of a user-object from a recommendation data set, and acquiring knowledge-graph information from a knowledge-graph data set; constructing an interaction matrix of the user-object according to the historical record of the user scoring object, and constructing a knowledge graph after mapping is established according to the knowledge graph data set and the entity of the recommendation data set; the user-article interaction information comprises a set of users and articles and a historical record of scoring articles by the users, and the knowledge graph consists of a head entity-relation-tail entity triplet;
step 2: inputting article characteristics obtained from the user-article interaction information and head entity characteristics in the knowledge graph to a multi-task learning sharing module;
step 3: training a knowledge graph module by an alternate training mode based on the multi-task learning sharing module, and obtaining a scoring function of the knowledge graph triplet by a similarity function;
step 4: and training a recommendation module by using an alternate training mode based on the multi-task learning sharing module, and finally obtaining the final prediction probability of the user participating in the project by using a prediction function.
2. The recommendation method of learning knowledge graph integrating interaction attentions according to claim 1, wherein the multitask learning sharing module comprises a plurality of feature attentions cross compression modules, the feature attentions cross compression modules comprise feature interaction attentions units and cross compression units, article features and head entity features are learned by the interaction attentions units to obtain article features and entity features integrating context information, and preparation is made for interactive training of the recommendation module and the knowledge graph module.
3. The recommendation method of learning knowledge graph with integrated interaction attention as in claim 1, wherein the recommendation module comprises a plurality of feature extraction layers and a prediction layer connected in sequence.
4. The recommendation method of learning knowledge graph with integrated interaction attention as set forth in claim 1, wherein the inputting the item features and the head entity features into the multi-task learning sharing module includes:
step 2.1: for each input vector of candidate articles and head entities, multiplying the vector by a trainable parameter matrix to obtain a query vector and an address vector of each vector;
step 2.2: calculating similarity scores of the query vector and the address vector, and then applying a normalized exponential function to the similarity scores to obtain respective weight coefficient scores;
step 2.3: weighting and summing the weight coefficient scores and the article vectors or the head entity vectors respectively to obtain article self-attention scores and head entity self-attention scores respectively;
step 2.4: performing point multiplication on the self-attention score of the article and the self-attention score of the head entity processed by the activation function respectively;
step 2.5: adding the point multiplication value and the initial feature vector to obtain an article vector and a head entity vector respectively;
step 2.6: the object vector and the head entity vector are subjected to characteristic crossing in a crossing compression layer to obtain a crossing multiplication matrix;
step 2.7: transposition is carried out on the cross multiplication matrix to obtain a transposition cross multiplication matrix;
step 2.8: and compressing the crossed features, and calculating and outputting the feature vectors of the articles and the head entities of the next layer.
5. The recommendation method of learning knowledge graph with integrated interaction attention as set forth in claim 4, wherein in step 2.2, normalized exponential functions are applied to the similarity scores to obtain respective weight coefficient scores expressed as:
wherein score1 and score2 are the weight coefficient score of the item and the weight coefficient score of the entity, respectively; d, d k Is the query vector Q 1 、Q 2 And address vector K 1 T 、K 2 T Vector dimensions of the matrix.
6. The recommendation method of learning knowledge graph with integrated interaction attention as claimed in claim 4, wherein in step 2.6, the feature intersection of the object feature vector and the head entity feature vector is performed in an intersection compression unit to obtain an intersection multiplication matrix, and the outer product of the object vector and the head entity vector is calculated to construct a d x d-dimensional intersection multiplication matrix, which is expressed as:
wherein ,BL ∈R dxd Is the layer L feature cross multiplication matrix, d is the dimension of the embedded vector,is B L The interaction of the item v with each feature of its associated header entity h.
7. The recommendation method of learning knowledge graph with integrated interaction attention as set forth in claim 4, wherein in step 2.8, the calculated output of the next layer of object and head entity feature vectors is expressed as:
wherein , and />The trainable compression unit weight vector and bias term for item vector v and head entity vector h at layer L are represented, respectively.
8. The recommendation method of learning knowledge graph with integrated interaction attention according to claim 1, wherein in step 3, the obtaining the scoring function of the knowledge graph triplet by the similarity function includes:
step 3.1: for a given triplet, the head entity obtains the high-order characteristic of the head entity through a multi-task learning sharing module;
step 3.2: extracting implicit cross characteristics of a relational entity by using a multi-layer neural network;
step 3.3: the head entity characteristics and the relation characteristics are embedded into a function through a knowledge graph to predict a prediction vector with a tail entity as a tail;
step 3.4: and obtaining a score function of the triplet through the similarity function.
9. The recommendation method of learning knowledge graph with integrated interaction attention according to claim 1, wherein in step 4, the training the recommendation module by the multi-task learning sharing module through the alternate training mode, and finally obtaining the final prediction probability of the user participation item through the prediction function comprises:
step 4.1: extracting compression characteristics of hidden layers of a user from the initial characteristic vector of the user through a plurality of neural network layers after characteristic embedding;
step 4.2: for the original input vector of the article, calculating to obtain the potential characteristics of the article through a multi-task learning sharing module;
step 4.3: and calculating a final prediction probability of the user participation project through the prediction function.
10. A recommendation system for learning knowledge maps with integrated interaction attention, comprising:
the data acquisition module is used for acquiring interaction information of the user and the object from the recommendation data set and acquiring knowledge-graph information from the knowledge-graph data set; constructing an interaction matrix of the user-object according to the historical record of the user scoring object, and constructing a knowledge graph after mapping is established according to the knowledge graph data set and the entity of the recommendation data set; the user-article interaction information comprises a set of users and articles and a historical record of scoring articles by the users, and the knowledge graph consists of a head entity-relation-tail entity triplet;
the interactive training preparation module is used for inputting the article characteristics and the head entity characteristics obtained from the interactive information of the user-article into the multi-task learning sharing module;
the knowledge graph training module is used for training the knowledge graph module in an alternating training mode based on the multi-task learning sharing module and obtaining a score function of the triplet through the similarity function;
and the recommendation module training module is used for training the recommendation module based on the multi-task learning sharing module in an alternating training mode, and finally obtaining the final prediction probability of the user participation project through the prediction function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310602037.0A CN116701611A (en) | 2023-05-25 | 2023-05-25 | Recommendation method and system for learning knowledge graph fusing interaction attention |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310602037.0A CN116701611A (en) | 2023-05-25 | 2023-05-25 | Recommendation method and system for learning knowledge graph fusing interaction attention |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116701611A true CN116701611A (en) | 2023-09-05 |
Family
ID=87840236
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310602037.0A Pending CN116701611A (en) | 2023-05-25 | 2023-05-25 | Recommendation method and system for learning knowledge graph fusing interaction attention |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116701611A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117370674A (en) * | 2023-12-08 | 2024-01-09 | 西南石油大学 | Multitask recommendation algorithm integrating user behaviors and knowledge patterns |
| CN117688247A (en) * | 2024-01-31 | 2024-03-12 | 云南大学 | Recommendation method, terminal device and storage medium |
-
2023
- 2023-05-25 CN CN202310602037.0A patent/CN116701611A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117370674A (en) * | 2023-12-08 | 2024-01-09 | 西南石油大学 | Multitask recommendation algorithm integrating user behaviors and knowledge patterns |
| CN117370674B (en) * | 2023-12-08 | 2024-02-09 | 西南石油大学 | Multitask recommendation algorithm integrating user behaviors and knowledge patterns |
| CN117688247A (en) * | 2024-01-31 | 2024-03-12 | 云南大学 | Recommendation method, terminal device and storage medium |
| CN117688247B (en) * | 2024-01-31 | 2024-04-12 | 云南大学 | Recommendation method, terminal device and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110083692B (en) | Text interactive matching method and device for financial knowledge question answering | |
| CN111897913B (en) | Semantic tree enhancement based cross-modal retrieval method for searching video from complex text | |
| CN110032635B (en) | Problem pair matching method and device based on depth feature fusion neural network | |
| CN110516085A (en) | The mutual search method of image text based on two-way attention | |
| CN116701611A (en) | Recommendation method and system for learning knowledge graph fusing interaction attention | |
| CN111414461B (en) | Intelligent question-answering method and system fusing knowledge base and user modeling | |
| EP4002161A1 (en) | Image retrieval method and apparatus, storage medium, and device | |
| Goh et al. | Food-image Classification Using Neural Network Model | |
| CN111753207B (en) | Collaborative filtering method for neural map based on comments | |
| CN111159485A (en) | Tail entity linking method, device, server and storage medium | |
| CN114693397A (en) | Multi-view multi-modal commodity recommendation method based on attention neural network | |
| CN115131698B (en) | Video attribute determining method, device, equipment and storage medium | |
| CN112016002A (en) | Mixed recommendation method integrating comment text level attention and time factors | |
| CN113177141A (en) | Multi-label video hash retrieval method and device based on semantic embedded soft similarity | |
| CN115687760A (en) | User learning interest label prediction method based on graph neural network | |
| CN116821291A (en) | Question-answering method and system based on knowledge graph embedding and language model alternate learning | |
| CN115471885A (en) | Action unit correlation learning method and device, electronic device and storage medium | |
| CN110502701B (en) | Friend recommendation method, system and storage medium introducing attention mechanism | |
| CN113409157B (en) | Cross-social network user alignment method and device | |
| CN111949894B (en) | Collaborative filtering personalized recommendation method based on multi-space interaction | |
| CN114398909A (en) | Question generation method, device, equipment and storage medium for dialogue training | |
| CN109284765A (en) | The scene image classification method of convolutional neural networks based on negative value feature | |
| CN115186072A (en) | Knowledge graph visual question-answering method based on double-process cognitive theory | |
| CN117131933A (en) | Multi-mode knowledge graph establishing method and application | |
| CN115619445A (en) | Fashion trend prediction method based on improved Transformer 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 |