CN114461907A - Knowledge graph-based multi-element environment perception recommendation method and system - Google Patents
Knowledge graph-based multi-element environment perception recommendation method and system Download PDFInfo
- Publication number
- CN114461907A CN114461907A CN202210079690.9A CN202210079690A CN114461907A CN 114461907 A CN114461907 A CN 114461907A CN 202210079690 A CN202210079690 A CN 202210079690A CN 114461907 A CN114461907 A CN 114461907A
- Authority
- CN
- China
- Prior art keywords
- user
- rule
- path
- item
- users
- 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
- 238000000034 method Methods 0.000 title claims abstract description 136
- 230000008447 perception Effects 0.000 title claims abstract description 31
- 230000007246 mechanism Effects 0.000 claims abstract description 27
- 238000004220 aggregation Methods 0.000 claims abstract description 25
- 239000013598 vector Substances 0.000 claims description 66
- 239000011159 matrix material Substances 0.000 claims description 46
- 238000012512 characterization method Methods 0.000 claims description 36
- 238000011176 pooling Methods 0.000 claims description 36
- 230000003993 interaction Effects 0.000 claims description 33
- 230000009466 transformation Effects 0.000 claims description 33
- 230000006870 function Effects 0.000 claims description 24
- 230000004931 aggregating effect Effects 0.000 claims description 22
- 238000001914 filtration Methods 0.000 claims description 17
- 230000002457 bidirectional effect Effects 0.000 claims description 12
- 238000000605 extraction Methods 0.000 claims description 12
- 238000006116 polymerization reaction Methods 0.000 claims description 11
- 238000013459 approach Methods 0.000 claims description 9
- 238000009826 distribution Methods 0.000 claims description 7
- 238000012545 processing Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 238000001514 detection method Methods 0.000 claims description 5
- 230000002452 interceptive effect Effects 0.000 claims description 5
- 230000007787 long-term memory Effects 0.000 claims description 5
- 238000013507 mapping Methods 0.000 claims description 5
- 230000015654 memory Effects 0.000 claims description 5
- 230000006403 short-term memory Effects 0.000 claims description 5
- 238000012935 Averaging Methods 0.000 claims description 4
- 230000002776 aggregation Effects 0.000 abstract description 24
- 230000014509 gene expression Effects 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 description 8
- 238000005070 sampling Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000006399 behavior Effects 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000005295 random walk Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 230000035945 sensitivity Effects 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000004132 cross linking Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000005065 mining Methods 0.000 description 2
- 238000013519 translation Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000012795 verification Methods 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic 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)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a knowledge graph-based multi-environment perception recommendation method and a knowledge graph-based multi-environment perception recommendation system. Then, a Bi-LSTM model is adopted to learn the semantic features of the paths generated by the rules, and an attention mechanism is adopted to distinguish the preferences of users for different rules and the rules which can represent the designated items most. Then, a local information sensing module is designed to learn the close-range neighborhood characteristics of the user and the project; and finally, integrating the user and item expressions learned by the global information perception module and the local information perception module through an aggregation algorithm, and implementing prediction. The experimental results on three real data sets prove that the performance of the method is obviously improved compared with that of a benchmark method.
Description
Technical Field
The invention belongs to the technical field of artificial intelligence, and particularly relates to a knowledge graph-based multi-element environment perception recommendation method and system.
Background
The recommendation system plays an important role in modern society, and has been widely applied to platforms such as online learning platforms, news websites, social media and online shopping. The core of the method is that the interest preference of a user is evaluated according to historical interaction information of clicking, browsing, watching, commenting and the like of the user on a platform, and knowledge, items or contents which may be interested in the user are recommended for the user. Among all recommendation methods, the most classical is Collaborative Filtering (CF), which finds and makes decisions for users mainly through behaviors among users (users) to find similarities between users or similarities between items (items). The collaborative filtering algorithm has become one of the most widely used algorithms in the recommendation system due to the advantages of universality, easy understanding and the like. However, the algorithm does not take the difference of user environments into consideration, and has more serious problems of data sparsity and cold start, and the performance of the algorithm has a larger space for improvement.
Today, the internet and platforms have rich user or project information available for use. Researchers began to consider introducing external knowledge as auxiliary information to enhance the user and item characterization at the time of recommendation to improve the performance of recommendation algorithms. Among all the auxiliary information, the structured knowledge base represented by the knowledge map is most concerned by researchers. The knowledge graph not only contains a large amount of high-quality structured data, but also can be conveniently read by a computer; and rich semantic association among entities in the knowledge graph can enhance the relation between the users and the items in the recommended scenes, and is beneficial to mining deeper preference relation between the users and the items.
Due to the natural high dimensionality and heterogeneity of the knowledge graph, how to effectively integrate the graph-based structural information such as the knowledge graph into the characterization of users or projects is a core problem to be solved by the knowledge graph-based recommendation system. To address this problem, existing solutions can be largely divided into 3 types: embedding-based methods, Path-based methods, Propagation-based methods.
The embedding-based method comprises the following steps: the method mainly comprises the steps of preprocessing by using a knowledge graph embedding algorithm to form low-dimensional vector representation of entities and relations, and then fusing the learned entity vectors into a recommendation system. Although the knowledge graph embedding-based method is applied to a recommendation system to a certain extent, the existing knowledge graph embedding method, including a translation distance model and a semantic matching model, focuses on modeling semantic association of entities in a knowledge graph, is more suitable for tasks such as knowledge graph completion and link prediction, and the learned embedding feature representation is poor in performance when applied to the recommendation system. And the knowledge graph embedding method is mainly used for embedding representation of a static graph, modeling and feature representation of the knowledge graph are independent from downstream tasks, and when a new user or item is added into a recommendation system, a knowledge graph embedding algorithm needs to be called to relearn feature representation of the whole graph.
The path-based method comprises the following steps: the method is also called a HIN (heterologous Information networks) -based method, and the method is used for recommending the project by constructing a user-project graph and then utilizing path connection between entities in the graph. Conventional path-based approaches typically combine the extracted paths in the HIN with mf (matrix factorization) methods and exploit the connection similarities of users and/or items to enhance recommendations. Path-based methods have a large improvement in both recommendation accuracy and interpretability compared to embedding-based methods, but such methods rely heavily on the quality of the Meta Path (Meta Path) employed, making it difficult in practice to optimize the selection of paths. In addition, the conventional method mainly adopts a manual mode for designing the meta-path, and is not suitable for the situations of more knowledge graph entities and more complex relation types.
The method based on propagation comprises the following steps: the method mainly integrates semantic representation of entities and relationships in an embedded-based method and semantic connections in a path-based method for recommendation. The multi-hop neighborhood information of the entity is aggregated by a mechanism that the embedded representation of the entity is embedded and transmitted on a knowledge graph, so that the preference and feature representation of users and items are enriched, and the recommendation effect is improved. Although the propagation-based method achieves better recommendation performance, the method also has some problems, which are highlighted in that: 1) the method is difficult to distinguish the correctness and the effectiveness of the neighborhood entity when representing the user and the project, and noise information is easy to introduce; 2) this type of approach does not explicitly embody the connection patterns between the user and his preferences as the path-based approach, making this type of approach less interpretable; 3) the propagation learning method represented by the graph neural network can learn that the neighborhood depth is usually less than or equal to 3, and deeper connection information between the user and the project cannot be acquired.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to provide a knowledge graph-based multi-environment perception recommendation method and system, and aims to solve the problem that reasonable meta-paths are difficult to manually create due to the fact that the knowledge graph entities and the relations are large in quantity and complex in type.
In order to achieve the above object, in a first aspect, the present invention provides a knowledge-graph-based multivariate environment perception recommendation method, including the following steps:
determining a path between the user and each project according to the interaction records generated between the user and the projects; extracting rules for generating paths between the users and the items from the paths according to the association among the plurality of items; the items are entities in a knowledge graph; the path is used for explaining how the user interacts with the item, and the rule is used for reflecting the interest preference of the user on the entity;
filtering the rules based on chi-square distribution according to the number of paths which are connected correctly and incorrectly by the rules and the number of paths which are not connected correctly and incorrectly by the rules, and keeping the high-quality rules;
determining all paths between the user and the project based on the high-quality rule, and eliminating noise paths in all the paths based on the weight of each path to obtain path samples between the user and the project;
learning the low-dimensional embedded representation of each path in the path sample by using a bidirectional long and short term memory network, and aggregating the low-dimensional embedded representation of the path corresponding to each rule to obtain long-distance multi-semantic connection information between the user and the project based on the global environment;
learning different attention scores of a plurality of neighborhood entities of the user and a plurality of neighborhood entities of the project based on a graph attention network, and aggregating the corresponding neighborhood entities through the different attention scores to respectively acquire close neighborhood information of the user and the project based on the local environment;
aggregating long-distance multi-element semantic connection information between users and projects, close-range neighborhood information of the users and close-range neighborhood information of the projects to obtain comprehensive characterization vectors of the users and the projects;
and inputting the comprehensive characterization vectors of the users and the items into a multilayer perceptron to obtain a prediction score between the users and the items, and recommending the corresponding items for the users based on the prediction score.
Optionally, according to interaction records generated between the user and a plurality of items, a bidirectional path search strategy is adopted to execute a breadth-first search strategy from the user and the items respectively to determine all connection paths between the user and the items, and the connection relation of each path is used as a rule;
given a rule, orderRepresenting all user-item pairs connected by interactive relations in the knowledge graph, u representing a user, and v representing an item;representing all user-item pairs in the knowledge-graph connected by the path generated by the rule; then:
tnnum=m*n-tpnum-fpnum-fnnumrepresenting the number of erroneous user-item pairs that are not connected by a rule; m represents the total number of users, and n represents the total number of items;
the obtained tpnum、fpnum、fnnum、tnnumForm a List table [ (tp)num,fpnum),(fnnum,tnnum)]Further, a chi-square detection method is adopted to carry out statistics on the list table to obtain a statistic value; if the statistical value is smaller than a set threshold value, the quality of the rule is high, and the rule is reserved; otherwise, the rule is removed.
Optionally, learning the embedded feature representation of each node in the knowledge graph by adopting a factor decomposition method; calculating the similarity between two adjacent nodes in the path generated by each rule by adopting cosine similarity based on the embedded characteristic representation of each node; averaging the similarity scores between two adjacent nodes in each path, and taking the average as the priority score of the path; finally, for each rule, the k paths with the highest priority are selected to represent the path samples between the user and the item with respect to the rule.
Optionally, the embedded representation of a given path of length s, p ═ eu,ew1,ew2,…,evUsing forward hidden state sequence of each entity on the Bi-LSTM learning pathAnd reverse hidden state sequenceThe mutual information of the learning node and the forward neighbor and the backward neighbor is used; e.g. of the typeuAnd evRepresenting users and items on the path, respectively, ew1、ew2… watchShowing other entities on the path;
connecting the forward and reverse state outputs of any unit in the bidirectional long-short term memory network (Bi-LSTM) as an entity e in the pathjFinal output h after Bi-LSTM treatmentj(ii) a Wherein e isj∈eu、ew1、ew2、…、ev;
For a path p with the length s, the output of the path p after the Bi-LSTM is a matrix [ h ] formed by the output states of all entitiesu,hw1,hw2,…,hv](ii) a Merging the matrixes by adopting a pooling method to obtain a final path embedded representation ep(ii) a Embedding paths into representation epOutput h processed by Bi-LSTM respectively with user and project nodeu、hvConnecting to obtain the representation e of the user and the project on the pathupAnd evp;
Aggregating all path-based user and item representations contained in each rule through a pooling method to obtain rule-based user and item embedded vector representation eurAnd evr;
Order toRepresenting rule-based user-embedded vector representation [ e ]ur1,eur2,…,eurl]A two-dimensional matrix is formed, wherein l is the number of set rules in the recommended scene, eurjFor rule j based users to embed vector representations, d is the length of each rule vector representation;
the regular attention coefficient score is obtained by the following formulaur:scoreur=LeakyReLU(W1Eur) (ii) a Wherein,is a first linear transformation weight matrix; LeakyReLU is an activation function;
normalizing the attention coefficients of all the rules by utilizing a softmax function to obtain the final attention coefficient of each ruleAttention weight vector: attenur=softmax(W2scoreur) (ii) a Wherein,representing a second linear transformation weight matrix; attenurRepresenting an attention weight vector;
finally, user u represents e based on rule embeddingur' is: e.g. of the typeur′=attenurEur;
Likewise, a rule-based embedded representation e of the item v can be obtainedvr′。
for user u, user u and its neighborhood are calculated respectivelySimilarity coefficient between: in the formula, W3Is the weight matrix of the third linear transformation, W4Is a weight matrix of a fourth linear transformation [ | | · [ ] | ]]Representation for user u and item viSplicing the linearly transformed features, and mapping the spliced high-dimensional features to a real number by using an a (-) function;
normalizing the attention coefficient by adopting a softmax function to obtain the normalized attention coefficient
Weighting and summing the characteristics of the neighborhood according to the normalized attention coefficient to obtain the characterization e of the user u aggregated with the local neighborhood characteristicsut:W5Is a weight matrix of a fifth linear transformation;
order toThe local neighborhood of the item v is represented, and the representation e of the item v with the local neighborhood characteristics aggregated can be obtained by using a graph attention mechanism as well as the processing of the user uvt。
Optionally, long-distance multi-semantic connection information e between the user and the project is connected in a connection aggregation modeur' and evr', user's close proximity information eutAnd close proximity information e of the itemvtPolymerizing to obtain a connected polymerization vector aggconcat:
aggconcat=σ(W8(eur′||eut||evr′||evt)+b)
Agg is preparedconcatAs a comprehensive characterization vector e for users and itemsuv;
E is to beuvInputting the data into a multi-layer perceptron to obtain the prediction scores of the users and the projects
In the formula, W8、W9A weight matrix representing the eighth linear transformation and a weight matrix representing the ninth linear transformation, and b represents a bias of the linear transformation.
In a second aspect, the present invention provides a knowledge-graph-based multivariate environment-aware recommendation system, comprising:
the rule-based path extraction module is used for determining a path between the user and each project according to an interaction record generated between the user and the projects; extracting rules for generating paths between the users and the items from the paths according to the association among the plurality of items; the items are entities in a knowledge graph; the path is used for explaining how the user interacts with the item, and the rule is used for reflecting the interest preference of the user on the entity; filtering the rules based on chi-square distribution according to the number of paths which are connected correctly and incorrectly by the rules and the number of paths which are not connected correctly and incorrectly by the rules, and keeping the high-quality rules; determining all paths between the user and the project based on the high-quality rule, and eliminating noise paths in all the paths based on the weight of each path to obtain path samples between the user and the project;
the global environment perception module is used for learning the low-dimensional embedded representation of each path in the path sample by utilizing a bidirectional long and short term memory network and aggregating the low-dimensional embedded representation of the path corresponding to each rule to obtain long-distance multi-element semantic connection information between the user and the project based on the global environment;
the local environment sensing module is used for learning different attention scores of a plurality of neighborhood entities of the user and a plurality of neighborhood entities of the project based on the graph attention network so as to aggregate the corresponding neighborhood entities through the different attention scores to respectively acquire close-range neighborhood information of the user and the project based on the local environment;
the scoring prediction module is used for aggregating long-distance multi-element semantic connection information between the users and the projects, close-range neighborhood information of the users and close-range neighborhood information of the projects to obtain comprehensive characterization vectors of the users and the projects; and inputting the comprehensive characterization vectors of the users and the items into a multilayer perceptron to obtain a prediction score between the users and the items, and recommending the corresponding items for the users based on the prediction score.
Optionally, the rule-based path extraction module is configured to respectively search the two-way path according to interaction records generated between the user and the plurality of items by using a two-way path search strategyStarting from the user and the project to execute a breadth-first search strategy to determine all connection paths between the user and the project, and taking the connection relation of each path as a rule; given a rule, orderRepresenting all user-item pairs connected by interactive relations in the knowledge graph, u representing a user, and v representing an item;representing all user-item pairs in the knowledge-graph connected by the path generated by the rule; then: indicating the number of correct user-item pairs that are connected by the rule;indicating the number of correct user-item pairs that are not connected by the rule;representing the number of erroneous user-item pairs that are connected by the rule; tnnum=m*n-tpnum-fpnum-fnnumRepresenting the number of erroneous user-item pairs that are not connected by a rule; m represents the total number of users, and n represents the total number of items; the obtained tpnum、fpnum、fnnum、fnnumForm a List table [ (tp)num,tpnum),(fnnum,fnnum)]Further, a chi-square detection method is adopted to carry out statistics on the list table to obtain a statistic value; if the statistical value is smaller than a set threshold value, the quality of the rule is high, and the rule is reserved; otherwise, the rule is removed.
Optionally, the obtaining, by the global environment sensing module, long-distance multivariate semantic connection information between the user and the project based on the global environment specifically includes:
given an embedded representation of a path of length s, p ═ eu,ew1,ew2,…,evUsing forward hidden state sequence of each entity on the Bi-LSTM learning pathAnd reverse hidden state sequenceThe mutual information of the learning node and the forward neighbor and the backward neighbor is used; e.g. of the typeuAnd evRepresenting users and items on the path, respectively, ew1、ew2… denote other entities on the path;
connecting the outputs of the forward and reverse states of any unit in the Bi-LSTM as an entity e in the pathjFinal output h after Bi-LSTM treatmentj(ii) a Wherein e isj∈eu、ew1、ew2、…、ev;
For a path p with the length s, the output of the path p after the Bi-LSTM is a matrix [ h ] formed by the output states of all entitiesu,hw1,hw2,…,hv](ii) a Merging the matrixes by adopting a pooling method to obtain a final path embedded representation ep(ii) a Embedding paths into representation epOutput h processed by Bi-LSTM respectively with user and project nodeu、hvConnecting to obtain the representation e of the user and the project on the pathupAnd evp;
Aggregating all path-based user and item representations contained in each rule through a pooling method to obtain rule-based user and item embedded vector representation eurAnd evr;
Order toRepresenting rule-based user-embedded vector representation [ e ]ur1,eur2,…,eurl]A two-dimensional matrix is formed, wherein l is the number of set rules in the recommended scene, eur jFor rule j based users to embed vector representations, d is the length of each rule vector representation;
the regular attention coefficient score is obtained by the following formulaur:scoreur=LeakyReLU(W1Eur) (ii) a Wherein,is a first linear transformation weight matrix; LeakyReLU is an activation function;
normalizing the attention coefficients of all the rules by utilizing a softmax function to obtain a final attention weight vector of each rule: atten (r)ur=softmax(W2scoreur) (ii) a Wherein,representing a second linear transformation weight matrix; attenurRepresenting an attention weight vector;
finally, user u represents e based on rule embeddingur' is: e.g. of the typeur′=attenurEur;
Likewise, a rule-based embedded representation e of the item v can be obtainedvr′。
Optionally, the local environment sensing module obtains near-distance neighborhood information of the user and the item based on the local environment, and specifically includes:
for user u, user u and its neighborhood are calculated respectivelySimilarity coefficient between: in the formula, W3Is the weight matrix of the third linear transformation, W4Is a weight matrix of a fourth linear transformation [ | | · [ ] | ]]Representation for user u and item viSplicing the linearly transformed features, and mapping the spliced high-dimensional features to a real number by using an a (-) function;
normalizing the attention coefficient by adopting a softmax function to obtain the normalized attention coefficient
Weighting and summing the characteristics of the neighborhood according to the normalized attention coefficient to obtain the characterization e of the user u aggregated with the local neighborhood characteristicsut:W5Is a weight matrix of a fifth linear transformation;
order toThe local neighborhood of the item v is represented, and the representation e of the item v with the local neighborhood characteristics aggregated can be obtained by using a graph attention mechanism as well as the processing of the user uvt。
Generally, compared with the prior art, the above technical solution conceived by the present invention has the following beneficial effects:
the invention provides a knowledge graph-based multi-element environment perception recommendation method and system, which combine the characteristics of a path-based method and a propagation-based method to improve the accuracy and diversity of recommendation results. Aiming at the problem that reasonable meta-paths are difficult to manually create due to the fact that the number of knowledge graph entities and relations is large and the types are complex, an automatic rule discovery algorithm based on user-project interaction relations is provided. The method comprises the steps of searching paths among all user-item pairs in a knowledge graph by mining user-item interaction records in a recommendation scene, and finding a common relation mode, namely a rule, in the paths. Then, a rule filtering method based on chi-square distribution is proposed, and the rules with high quality are found out from all the discovered paths, and not only represent specific structural and semantic features in the knowledge graph, but also are the most representative user preference modes in the recommended scene because the rules come from all the user-item interaction records. More importantly, compared with the existing path-based or propagation-based method, when the user and item feature representation is learned, different meta-modes are respectively adopted to respectively represent the user and the item, and the problem of information association between the user and the item in the knowledge graph is ignored.
The invention provides a knowledge graph-based multi-environment perception recommendation method and a knowledge graph-based multi-environment perception recommendation system, aiming at a path-based method, when the user and project feature characterization is carried out, a sampling mode (mainly a random walk method based on a meta path) is mainly adopted to generate feature paths of users and projects, and the path-based mode only uses long-distance feature information between the users and the projects and ignores local neighborhood information of the users and the projects. In order to make up for the defect of the path-based mode in user-item feature learning, the invention designs a user and item feature representation method based on local neighborhood information. Inspired by a collaborative filtering algorithm based on users and a collaborative filtering algorithm based on projects, when local neighborhood information is learned, the characteristics of a target user are mainly represented by the projects which generate interactive relations with the target user; the characteristics of the target item are represented by the user with whom the interaction is made. And learning the local environmental characteristics of the user and the project by adopting the attention network.
The invention provides a knowledge graph-based multi-element environment perception recommendation method and system, and provides an automatic rule discovery method which can automatically discover the most representative user preference mode in a recommended scene according to a given knowledge graph and user behaviors in the recommended scene, so that the problem that the conventional path-based recommendation algorithm needs to manually design an element path is solved. The invention provides a new recommendation model MANN, which simultaneously learns the high-order semantic relation and the local neighborhood characteristics of a user-item in a knowledge graph, and adopts an attention mechanism to perform characteristic fusion, thereby realizing high-quality user-item characteristic representation. The experimental results of three real data sets of film, book and news show that the method has obvious advantages compared with the existing recommendation algorithm method based on the knowledge map.
Drawings
FIG. 1 is a flowchart of a knowledge-graph-based multivariate environment-aware recommendation method provided by an embodiment of the present invention;
FIG. 2 is a block diagram of the overall framework of the MANN model provided by an embodiment of the present invention;
FIG. 3 is a diagram of a knowledge-graph based user-item interaction provided by an embodiment of the present invention;
FIG. 4 is a Bi-LSTM model-based path coding diagram provided by an embodiment of the present invention;
FIG. 5 is a diagram illustrating an example of a local neighborhood representation of users and items provided by an embodiment of the present invention;
FIG. 6 is a graph of the impact of different pooling methods provided by embodiments of the present invention on a MANN;
FIG. 7 is a diagram illustrating the impact of various information aggregation methods provided by embodiments of the present invention on a MANN;
FIG. 8 is a diagram of a knowledge-graph based multivariate environment-aware recommendation system architecture provided by an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the invention.
A knowledge-graph-based recommendation system refers to learning feature representations of users and items by introducing information of a knowledge graph into a recommendation scene and then generating recommendations by using the features. In a typical recommendation scenario, m users U ═ U are included1,u2,…,umAnd n items V ═ V }1,v2,…,vn}. Obtaining a user-item interaction matrix through user implicit feedback information contained in the log record, such as clicking, watching, commenting and the likeIf user u has interacted with item v, then y uv1 is ═ 1; otherwise, yuv0. In addition, y isuv0 does not indicate that user u dislikes item v, which may be because user u ignores item v.
For a knowledge graph, it can be expressed asWherein the triple (h, r, t) represents the existence of the relation r, epsilon and t between the head entity h and the tail entity t in the knowledge graphRepresenting a set of entities and a set of relationships, respectively. For example, a triple (school and society, author, john duckwey) indicates the fact that "john duckwey writes a book" school and society ". In many recommendation scenarios, the item at the time of recommendation is related to an entity in the knowledge-graph. For example, in the recommendation of learning resources, a book "school and society" may be an entity in the knowledge graph. Additionally, the set A { (V, w) | V ∈ V, w ∈ ε } is introduced to represent the alignment of the item with the entity in the knowledge-graph, where the concept pair (V, w) represents that the item V can be aligned to the entity w in the knowledge-graph.
To link a user set U with an item set V toKnowledge graph G1We introduce a new concept called the collaborative knowledge graph. First, each user behavior is represented as a triple (u, interaction, v); all user behavior triplets are then integrated into the knowledge-graph G based on the item-entity alignment set A1In the method, a unified knowledge graph is formed Wherein ε' ═ ε @ U,in addition, in G, useAndto represent low-dimensional embedded representations of entities, relationships, users, and items.
Rules (Rules) refer to a set of relationships generated from a derived path of relationship interactions that account for the reason that a user interacts with an item. Let Rule be { Rule1,rule2,…,rulelDenotes a Rule set, where any Rule of Rule is represented by a series of relationships { p }1,p2,…,phIs composed of(symbol)P on the left is the rule header, p on the right1∧p2∧…∧phCalled a regular body, the number of which is the path length. When user u and item v instantiate rule, this means that user u and item v are represented by entity set { w1,w2,…,wh-1In the path Wherein (u, r, v) ∈ G.
According to the method, a user-item interaction matrix Y and a knowledge graph G are given, and the goal of a recommendation algorithm is to predict whether a user u has potential interest in an item v. The formalized representation is: learning functionWhereinRepresents the probability that user u will interact with item v, with Θ being a functionThe parameter (c) of (c).
Fig. 1 is a flowchart of a knowledge-graph-based multivariate environment-aware recommendation method according to an embodiment of the present invention, as shown in fig. 1, including the following steps:
s101, determining a path between a user and each project according to interaction records generated between the user and a plurality of projects; extracting rules for generating paths between the users and the items from the paths according to the association among the plurality of items; the items are entities in a knowledge graph; the path is used for explaining how the user interacts with the item, and the rule is used for reflecting the interest preference of the user on the entity;
s102, filtering the rules based on chi-square distribution according to the number of paths which are connected correctly and incorrectly by the rules and the number of paths which are not connected correctly and incorrectly by the rules, and keeping the high-quality rules;
s103, determining all paths between the user and the project based on the high-quality rule, and eliminating noise paths in all paths based on the weight of each path to obtain path samples between the user and the project;
s104, learning the low-dimensional embedded representation of each path in the path sample by using a bidirectional long-short term memory network, and aggregating the low-dimensional embedded representation of the path corresponding to each rule to obtain long-distance multi-element semantic connection information between the user and the project based on the global environment;
s105, learning different attention scores of a plurality of neighborhood entities of the user and a plurality of neighborhood entities of the project based on the graph attention network, and aggregating the corresponding neighborhood entities through the different attention scores to respectively obtain close-range neighborhood information of the user and the project based on the local environment;
s106, aggregating long-distance multi-element semantic connection information between the users and the projects, close-range neighborhood information of the users and close-range neighborhood information of the projects to obtain comprehensive characterization vectors of the users and the projects;
s107, inputting the comprehensive characterization vectors of the users and the items into a multilayer perceptron to obtain prediction scores between the users and the items, and recommending the corresponding items for the users based on the prediction scores.
As shown in fig. 2, the multivariate environment-aware recommendation algorithm MANN mainly includes four modules, which are: rule-based path extraction, global context awareness, local context awareness, and score prediction. The rule extraction and path generation module is mainly responsible for processing rule extraction of the knowledge graph and path sampling based on the rule; the rule-based global environment perception module mainly processes the characteristic representation of users and items based on paths; the local environment perception module is mainly used for learning local neighborhood characteristics of users and projects. The scoring prediction module is mainly used for aggregating the global environment perception characteristics and the local environment perception characteristics and predicting scoring between the user and the project.
In a specific embodiment, the rule-based path extraction provided in the embodiment of the present invention specifically includes the following steps:
the purpose of rule extraction is to find the best quality relationship sequence from all the user-item interaction pairs based on entity connection in the knowledge graph so as to represent the reason for the interest of the user in the item. The rule extraction method mainly comprises three steps of rule discovery, rule filtering and path sampling.
1) Rule discovery (rule finding)
As can be appreciated from the definition of the above rules, the candidate rules are all sequences of relationships in G in other paths connecting two entities in a triplet (u, interaction, v). Taking FIG. 3 as an example, if user u has interaction records with the movie "Raiders of the Lost Ark", all connection paths between the two can be considered as candidate rules. These rules explain why user u likes the movie "Raiders of the Lost Ark". For example, the slave pathIt can be concluded that one possible reason that user u likes the movie "Raiders of the Lost Ark" is that the movie is "adventure" with the genre of the movie "Cast Away". Assuming that user u has interacted with multiple items, after finding paths between all user-item pairs and finding rules in those paths, the user's interest preferences may be characterized by those rules. By analogy, in a recommendation scenario of m users and n items, after discovering high-quality rules among all triples (u, interaction, v), the preferences of the users can be defined through the rules, so as to realize recommendation.
Because the knowledge graph comprises a plurality of entities and relations, in order to accelerate the rule discovery speed, the MANN algorithm adopts a two-way path search strategy. For each triplet (u, interaction, v) contained in G, a breadth-first search strategy is executed starting from nodes u and v respectively to discover all connection paths between nodes u and v, and the connection relationship in each discovered path is taken as a rule. During the search process, the MANN algorithm maps G knowledgeYTreated as an undirected graph and set the maximum path length to s.
In fig. 2, from the triplet (u, interaction, v), the following three rules are available:
2) rule filtering (rule filtering)
Because knowledge maps contain complex concepts and relationships, the discovered rules are of varying quality. With rule r generated in fig. 2:for example, the interaction between users u and v may generate a path through the rule:
this path illustrates that the reason why user u likes the movie Raiders of the Lost Ark is that the director of the movie Raiders of the Lost Ark is Steven Spielberg, and Steven Spielberg has had collaboration with Tom Hanks, who has participated in the movie Cast Away. It is clear that the path generated by this rule is not very interpreted and can be considered as noise information. Therefore, to improve the performance of the recommendation algorithm, it is necessary to filter out low quality rules.
The MANN algorithm has designed two methods to filter the low quality rules: a rule confidence based approach.
Rule confidence based method: given a rule, orderRepresenting all user-item pairs in G that are interactively connected by a relationship;representing all user-item pairs in G that are connected by the rule-generated path. It is possible to obtain:
tnnum=m*n-tpnum-fpnum-fnnumindicating the number of erroneous user-item pairs that are not connected by a regular rule.
The obtained tpnum、fpnum、fnnum、tnnumForm a List table [ (tp)num,fpnum),(fnnum,tnnum)]And further carrying out statistics on the list table by adopting a chi-square detection method to obtain a p value. If the p value is smaller than the set threshold value thetachi(0.05), the rule is indicated to have higher quality, and the rule is reserved; otherwise, the rule is removed.
3) Path sampling (rule sampling)
Existing meta-path or metagraph based path sampling mainly uses two types of methods. The first method is to design different metagraph (Meta-graph) patterns and follow these patterns to extract metagraph instances, i.e. paths, using a random walk algorithm; the second method first learns the embedded vector representation of each node from the metagraph by using a Factorization Machine (FM) and a Bayesian Personalized Ranking (BPR); then, based on the embedded vectors, calculating the similarity between the neighborhood nodes in the metagraph; then, the average value of the similarity between any adjacent nodes in the path is used for representing the priority of the path; and finally, sorting the paths according to the priorities of the paths, and then selecting Top-K paths as final output. Although the two methods are widely applied, the two methods have some problems. For the first method, the path instances generated by the random walk strategy used in the method are generally low in quality and contain much noise information. For the second category of methods, the FM and BPR algorithms used by them cannot learn the interaction information between the user, the item and its attributes. More importantly, although the existing two types of methods can respectively depict the structure and semantic information of the user and the item in the knowledge graph through the meta path or the meta graph, the connection relationship of the user and the item in the knowledge graph is ignored, and the relationship is the most important characteristic for describing the user and the item in the recommendation scene.
The invention provides a new rule-based path sampling method. Since the rule in the invention is generated according to the interaction relationship between the user and the item, from the user perspective, the rule represents a type of user preference mode; from an item perspective, the reason why the rule is liked by the user is indicated. Therefore, feature learning is performed on the paths generated by different rules, and feature representation of the user and the project can be enriched.
For a given user-item pair (u, v) and Rule set Rule ═ Rule1,rule2,…,rulelAnd finding all paths connecting the user u and the item v based on each rule, wherein the paths are expressed asWherein Representing all paths p generated based on ruleruleA collection of (a). Since the knowledge-graph contains complex entities and relationships, the generated path may contain noise or the path itself may be a noisy path. To this end, the MANN model employs a path weight representation to eliminate the noise path. Firstly, learning the embedding characteristics of each node in G by adopting a factorization machine methodSymbolizing; then, calculating the similarity between two adjacent nodes in the path generated by each rule by adopting cos similarity; averaging the similarity scores between two adjacent nodes in each path, and taking the average as the priority score of the path; finally, for each rule, the k paths with the highest priority are selected to represent the path samples between user u and item v for that rule, i.e.
In a specific embodiment, the rule-based global context awareness provided by the embodiment of the present invention specifically includes the following components:
1) Bi-LSTM based path-embedded representation
A path is composed of a sequence of entities, and there is an ordered relationship between the entities in the sequence. One widely used method of sequence representation is to treat the path as a sentence, treat the nodes in the path as words in the sentence, and then use the Word2Vec model to learn the embedded vector representation of each Word. Although research has proven that this approach is effective in common networks; however, in a network with node attributes, the node attributes affect each other, and the Word2Vec model cannot learn such auxiliary information. Another type of work is to integrate discovered paths into deep learning models using Graph Convolutional Networks (GCNs), but the aggregation operations in GCNs do not take into account the order relationship of the nodes.
In order to solve the problems of the path algorithm, the invention provides a Bi-LSTM-based path embedded representation model, as shown in FIG. 4. Because each path is a sequence formed by entities/nodes, the nodes and the neighbors thereof are connected by a certain type of relation in rule; if the position of a node in a path changes, the semantic information generated by the path will change accordingly. The LSTM model can effectively capture the sequence relationship between nodes and the long-distance dependency between nodes, and has been widely used in the field of natural language processing. To do this, we use the LSTM model to learn the low-dimensional embedded representation of each path. As described in the previous section, each path connects a user-item pair, which expresses the relationship (u, interaction, v), that is, each path not only conveys the interest preference of user u, but also characterizes item v. Because the Bi-LSTM model considers information in both the front and back nodes, the Bi-LSTM model is used to model paths for this purpose.
Given an embedded representation of a sequence of entities (i.e., a path) of length s, p ═ eu,ew1,ew2,…,evUsing Bi-LSTM to learn forward hidden state sequencesAnd reverse hidden state sequenceEach hidden state can be regarded as a node, and the Bi-LSTM model can learn the mutual information of the node and forward neighbors and backward neighbors. In addition, the corresponding node attribute is stored in each hidden state, thereby enabling a good fusion of network characteristics and node attribute information. The Bi-LSTM mainly comprises an Input gate (Input gate), a forgetting gate (Forget gate), a Cell state (Cell state), an Output gate (Output gate) and other modules, and is formed by a unit in a forward hidden state sequenceThe specific processing procedures are described as follows:
let ejPresentation unitThe corresponding entity represents. The input gate includes a weight matrix WxiAnd WhiAnd an offset vector bi:
The forgetting gate comprises a weight matrix WxfAnd WhfAnd an offset vector bf:
The cell states include a weight matrix WxcAnd WhcAnd an offset vector bc:
The cell state receives information from the hidden stateIf node j has only one predecessor node, thenIs the leading cell state ck(ii) a If it has multiple front-end nodes, thenThe value of (c) is the average Pooling (Avg-Pooling) value of all front node cell states:
then, the hidden state of node j can be obtained:
the output gate includes a weight matrix WxoAnd WhoAnd an offset vector bo:
in the above formulaDenotes the Hadamard product, tanh denotes the hyperbolic tangent function, and σ denotes the sigmoid function.
Since the Bi-LSTM contains both forward and backward information streams, for any unit in the Bi-LSTM, its forward and backward state outputs are connected (concat) as entity e in the pathjFinal output after Bi-LSTM treatment:
for a path p with the length s, the output of the path p after the Bi-LSTM is a matrix [ h ] formed by the output states of all entitiesu,hw1,hw2,…,hv](ii) a Then a pooling (Pooling) method is applied to merge the matrices [ h ]u,hw1,hw2,…,hv]To obtain a final path-embedded representation ep=pooling([hu,hw1,hw2,…,hv]). Embedding representation e into a pathpAnd connecting the Bi-LSTM output of the user node and the project node to obtain the representation of the final two nodes of the user node and the project node on the path:
eup=[hu||ep] (9)
evp=[hv||ep] (10)
2) rule Embedded representation (rule embedding)
Any rule contains k paths, and through the steps, embedded representations of users and items on the paths are obtained. Now, by aggregating the rules involvedAll path-based user and item representations are obtained to obtain rule-based user and item embedding vector representation eurAnd evr. Here, the invention uses a pooling approach to aggregate the paths:
3) multi-rule aggregation based on attention mechanism
Attention mechanisms stem from the study of human vision. In cognitive science, humans selectively focus on a portion of all information while ignoring other visible information due to bottlenecks in information processing. Based on the thought, the attention mechanism is introduced into the neural network, so that the system can learn to ignore irrelevant information and focus on important information, further the understanding of how the information in the input sequence influences the final output is facilitated, and the internal mechanism of the model is better understood. Attention mechanisms have been applied to many areas today, such as image recognition, machine translation, and recommendation systems.
Each rule generated by the method embodies a preference style. A variety of rules are typically included in a recommendation scenario to reflect different types of preferences or characteristics. While different rules contribute differently to the characterization of the characteristics of the user or item. For example, a user may like a certain genre of movies, and the user's attention will be focused primarily on movies of that particular genre. If the user likes a movie with a certain designated actor, the rules representing the semantics of the particular actor are more focused. Therefore, to distinguish the importance of different rules to user preferences and item characterization, a mechanism of attention is employed to learn the impact of different rules on users and items.
Order toRepresenting based on a gaugeThe user of [ e ] is embedded in the representationur1,eur2,…,eurl]And forming a two-dimensional matrix, wherein l is the number of the set rules in the recommended scene. The regular attention coefficient score is obtained by the following formulaur:
scoreur=LeakyReLU(W1Eur) (13)
Wherein,is a linear transformation weight matrix; LeakyReLU is an activation function that controls the value of the angle a of the negative slope to be 0.2.
Normalizing the attention scores of all the rules by utilizing a softmax function to obtain the final weight of each rule:
attenur=softmax(W2scoreur) (14)
wherein,representing a linear transformation weight matrix; attenurAn attention weight vector is represented.
Finally, user u represents e based on rule embeddingur' is:
eur′=attenurEur (15)
likewise, a rule-based embedded representation e of the item v can be obtainedvr′。
In a specific embodiment, the neighborhood local environment sensing provided by the embodiment of the present invention specifically includes the following parts:
the rule-based global environment sensing module acquires long-distance multi-element semantic connection information between the user and the project by adopting a path-based method, and ignores short-distance neighborhood information of the user and the project. The close proximity information is an important element representing the characteristics of the user and the item itself. Generally, items with which a user has interacted in a history can represent the user's preferences to some extent; likewise, the characteristics of all users who have interacted with an item can also represent the item to some extent, as shown in FIG. 5. In order to enrich the characteristic representation of users and projects, the invention provides a method for aggregating the neighborhood information of knowledge graph entities based on a graph attention machine mechanism. The graph attention mechanism may compute different attention scores for multiple neighborhood entities of a user/item. The MANN model aggregates neighborhood entity embedding by different attention scores to represent a local context-aware representation of a user or item.
Order toRepresenting a local neighborhood of user u. First, for user u, user u and its neighborhood are calculated respectivelySimilarity coefficient between:
in the formula, W3、Is a linearly transformed weight matrix, [ | | | - ]]Representation for user u and item viAnd (3) splicing the linearly transformed features, and mapping the spliced high-dimensional features onto a real number by using the a (-) function.
The attention coefficient was then normalized using the softmax function:
finally, the feature of the neighborhood is weighted and summed according to the attention coefficient to obtain the representation e of the user u aggregated with the local neighborhood featureut:
W5Is a weight matrix of linear transformation, such thatThe local neighborhood of the item v is represented, and the representation e of the item v with the local neighborhood characteristics aggregated can be obtained by using a graph attention mechanism as well as the processing of the user uvt。
In a specific embodiment, after two modules of global environment perception and local environment perception, a user and item representation e based on global environment perception is obtainedur' and evr', and user and item characterization e based on local environmental perceptionutAnd evt. For the polymerization of these four features Γ ═ eur′,eut,evr′,evtThe MANN model uses three information aggregation methods to obtain the comprehensive characterization e of users and itemsuv。
Sum aggregators (Sum aggregators). Summation aggregation is mainly the addition of the vectors of each vector, and then the linear transformation of the resultant vector:
pool polymerisation (Pooling agglomerator). The pooling polymerization mainly takes the maximum value on the same element of each vector, and then the resultant vector is subjected to linear transformation:
aggpool=σ(W7poolmax(Γ)+b) (20)
coupling polymerization (Concat agglomeror). The connection aggregation is mainly to splice the vectors in different directions, and then to perform linear transformation on the spliced vectors:
aggconcat=σ(W8(eur′||eut||evr′||evt)+b) (21)
obtaining a comprehensive characterization e of users and itemsuvAfter that, e is mixeduvInputting the data into a multi-layer perceptron to obtain a user-item scoring prediction:
W6、W7、W8and W9Is a weight matrix of the linear transformation. ReLU is used as the activation function because it better captures relevant features and fits training data.
To balance the number of positive and negative examples and to guarantee the training effect of the model, we extract the same number of negative examples as the positive examples for each user. The loss function used is represented as follows:
in the formulaRepresenting the cross entropy loss. The theta indicates the parameters of the model,indicating L2 regularization.
To evaluate the effectiveness and versatility of MANNs, we performed recommendation experiments in different scenarios: music, books, movies. In addition to the domain differences, the size and sparsity of the three data sets are also different.
FM, last: fm online music system the data set contains music listening information from more than 2000 users of the last.
Book-crosslinking: the data set contains scores (from 0 to 10) for over 100 million books from the Book-crosslinking website.
MovieLens-1M: the data set contains about 6000 movie viewing records of more active users.
Fm, boost-cross and MovieLens-1M are explicit feedback, so they need to be converted into implicit feedback. Fm and boost-cross do not set any threshold for positive feedback in movilens-1M dataset due to sparsity of data. As for negative samples, for each user, a record equal to the number of its positive samples is randomly drawn from the items it interacts, denoted as negative feedback.
TABLE 1 recommendation scenario oriented data set summarization
In the construction of the knowledge graph, domain-related subgraphs were extracted from the Satori6 knowledge graph from Microsoft, similar to the research and research, for the MovieLens-1M, Book-Cross and last. Items that match multiple entities and items that do not match any entity are then deleted from the subgraph. Table 1 summarizes detailed statistics for three data sets: fm (music), Book-cross, and MovieLens-1M (movie).
To prove the effectiveness of the MANN algorithm proposed by the present invention, MANN was compared with 8 recommendation methods:
SVD: the method is a classical collaborative filtering based recommendation method that uses inner products to model user-item interactions.
And LibFM: the method uses mainly a feature decomposition method to predict click rate, which uses the original features of the user and the project as input.
LibFM + TransE: the method extends the LibFM model by appending a transit learned entity representation to each user-item pair.
CKE: the method is based on an embedded recommendation representative method, which combines a collaborative filtering module with information of the structure, text, vision, etc. of the project in a unified Bayesian framework. In the experiment, CKE is realized as a framework combining a collaborative filtering module and a structured knowledge module.
PER: the method is a path-based recommendation representative method which treats a knowledge graph as a heterogeneous information network and extracts potential features based on meta paths to represent the connection relation between users and items.
RippleNet: the method is a propagation-based method that uses a memory network-like approach to propagate user preferences across a knowledge graph to implement recommendations.
KGCN: the method is a propagation-based method that selectively aggregates neighborhood information through an attention mechanism and applies graph-convolved neural networks to the knowledge-graph to learn the structure and semantic information of the knowledge-graph, as well as the personalized features and potential interests of the user.
KGAT: the method is a propagation-based method, and mainly utilizes an attention mechanism to distinguish the importance of entity neighborhoods in the information propagation process.
In the experiment, the proportion of the training set, the verification set and the test set is 6:2:2, and the final experiment results are obtained by averaging the five experiment results. The experimental situation is click rate prediction, and the performance of the model is measured by an AUC value and an F1 value. In the aspect of optimization, Adam is used for learning model parameters, and Xavier is used for initializing the model parameters. The MANN model is implemented by a Pythrch deep learning framework.
In the aspect of setting of the hyper-parameters, embedding a vector dimension d epsilon {8,16,32,64,128 }; the hidden layer dimension of the Bi-LSTM model is 2 x d; the size of Batch is set to 128, 1024 and 64 in three data sets of music, books and movies, respectively; learning rate eta ∈ {10 ∈ }-4,5*10-4,10-3,5*10-3,10-2B, }; regularization parameter λ ∈ {10 ∈ }-6,5*10-6,10-5,5*10-5,10-4,5*10-4,10-3,5*10-3,10-2B, }; the regular length is less than 5; the path quantity l of each rule sample belongs to {5,10 }; and the local neighborhood size t of the user and the item is larger than the size t of the user and the item, wherein the size t belongs to {8,12,16 and 32 }.
The experimental results of the MANN and the 8 recommended methods on the three data sets are shown in table 2. Overall, compared to the prior art, the experimental results of MANN were greatly improved in the three data sets, and the AUC indicator (area under the working characteristic curve of the subject) values of MANN were respectively improved (0.007,0.013,0.006) and the F1 indicator (equilibrium mean) values (0.026,0.014,0.011) in the three data sets compared to the prior art KGAT method with the best performance. In addition, MANN performed worse than RippleNet (-0.005) on AUC values for the data set MovieLens-1M, but better than RippleNet on F1.
In addition, as can be seen from table 2, the three recommended methods of SVD, LibFM, and LibFM + TransE, which do not use auxiliary information, are inferior to the other methods in both evaluation indexes of the three data sets. This also illustrates that the introduction of auxiliary information can play a positive role for the recommendation algorithm. In addition, the results of the recommendation algorithms MANN, KGAT and RippleNet using the knowledge graph are far better than those of other algorithms, which shows that the knowledge graph can play a great positive role in improving the performance of the recommendation system.
The MANN method is better than the three methods of RippleNet, KGCN and KGAT, and the reason is probably that the MANN adopts a more complex model structure to represent the interaction relationship between the user and the project; in addition, compared with a KGCN, KGAT and a RippleNet random sampling mode, the path generation algorithm adopted by the MANN can better learn high-quality semantic and structural information in a knowledge graph; and the MANN considers different interest preferences represented by different rules, adopts a plurality of Bi-LSTM algorithms to learn semantic information represented by the different rules, and uses an attention mechanism to depict personalized preferences of users and feature representation of items.
TABLE 2 results of AUC and F1 in click-through prediction
Specifically, the method provided by the present invention was tested in several ways:
1) sensitivity test of feature embedding characterization dimension d
The feature embedding characterization dimension determines the ability of the MANN model to characterize the feature. If the dimension is set too high, it will introduce noise information, resulting in under-fitting of the model; if the dimension is set too low, it does not perform well enough, resulting in an overfitting of the model. To investigate the effect of feature embedding characterization size on the model, experiments were performed with embedding sizes for all three datasets set in the {8,16,32,64,128} range.
Table 3 results of AUC and F1 in dataset music for different dimensions
Table 4 results of AUC and F1 in data set books for different dimensions
TABLE 5 results of AUC and F1 in a dataset movie for different dimensions
The results of the experiments on the MANN for the different dimensions are shown in tables 3-5, corresponding to the data sets music, books and movies, respectively. The model, as a whole, initially shows an increasing trend and then begins to decline in performance. The embedded characterization dimensions for which the best performance is found are different for each data set. For music and movie data sets, the optimal embedding characterization dimension is 32, whereas for books the optimal embedding characterization dimension is 16. The possible reason for this is that the more user interactions with features in the dataset, the more suitable the MANN model is for high embedded characterization dimensions to better characterize the feature information of each node on the path.
2) Sensitivity testing of pooling by different pooling methods
Pooling is a technique for aggregating features that can remove redundant information, reduce parameters of the neural network, and prevent model overfitting. After the characteristics of each node on the path are acquired through the Bi-LSTM, the MANN model adopts a pooling mode to aggregate the characteristics of the nodes to form a path representation; in addition, after obtaining the representation of the paths on each rule, the MANN model aggregates the characteristics of the paths in a pooling manner to form the representation of the rules. The impact of the maximum pooling method and the average pooling method on the MANN model will be discussed in this section.
The results of the different pooling methods on MANN are shown in FIG. 6. As can be seen from the table, the maximal pooling method (max-pooling) gives better experimental results than the average pooling method (ave-pooling) on all datasets and evaluation indices. The possible reason for this is that the max-pooling method can learn the most expressive information in each feature dimension when performing information aggregation. While the average pooling method considers all the feature information, it also loses the characterization power of each dimension on the semantic information.
3) Effectiveness of attention mechanism
Because the rules represent different types of semantic and structural information, the MANN model uses an attention mechanism to aggregate the embedded feature vectors of the rules to form a feature representation of the user and the project. In order to verify the influence of an attention mechanism on a MANN model, a maximum pooling method is adopted to replace the attention mechanism to aggregate feature representations of all rules, then the feature representations are respectively spliced with an initial user embedded vector representation and an initial project embedded vector representation, and a global environment perception representation of a user and a project is formed through two different linear layers. Here, the MANN model with the attention mechanism removed is named MANNw/o Atten。
Table 6 shows KGAT and MANNw/o AttenAnd the results of the three models of MANN on two metrics, AUC and F1. As can be seen, the MANNw/o AttenThis demonstrates that the attention mechanism plays a crucial role in the rule-based characterization of users and items, which effectively identifies the importance of different rules in representing user and item preferences.
TABLE 6 Effect of attention mechanism on MANN
4) Availability of local context awareness module
In addition to using a rule-based global context awareness module to learn user-item features, the MANN model introduces a local context awareness module. To verify the impact of the local context awareness modules on the MANN model, we removed the local context awareness modules from the MANN model and only used the global information awareness modules to learn the characteristics of users and projects. Here, the MANN model with the local environment awareness module removed is named MANNw/o local-context。
Table 7 shows the MANNw/o local-contextAnd the results of the two models of MANN on the three datasets. As can be seen, the MANNw/o local-contextAll metrics for the three datasets were worse than those of the MANN, indicating that the local context awareness module is functioning in the characterization of users and items. One possible explanation is that the rule-based global context awareness module only learns long-distance connection information between users and projects, but ignores local neighborhood information for both users and projects, while the local context awareness module makes up for the shortcomings of the global context awareness module.
TABLE 7 influence of local context awareness modules on MANN
5) Sensitivity testing of different information aggregation methods
The MANN model uses 3 information aggregation methods in the prediction phase: sum aggregation, pooling aggregation, and join aggregation. Different aggregation methods embody different integration methods for global environment perception information and local environment perception information of users and projects. For this purpose, experiments were carried out to study the effect of different polymerization methods on the MANN model.
The experimental results of the different information aggregation methods on MANNs are shown in fig. 7. It can be seen from the figure that regardless of which data set and evaluation index, the join aggregation performs better than the sum aggregation and pooling aggregation. The possible reason for this is that the join aggregation can retain more information contained in the embedded feature vector than the other two aggregation methods. In addition, it can be seen from the figure that the results of the sum polymerization and the pooling polymerization are almost similar, and the difference between the performances of the sum polymerization and the pooling polymerization is larger than that of the join polymerization; the possible reason is that the global environment perception and the local environment perception of the users and the items generate four feature vectors, each dimension of the four feature vectors represents different features, the summing aggregation and the pooling aggregation are the summing and maximization operation on each dimension of the four vectors, and hidden feature information of each dimension is not aligned.
The invention provides a knowledge graph-based multi-element environment perception recommendation model for resource recommendation. Firstly, a recommendation scene-oriented rule extraction method based on a knowledge graph is designed, and the method mainly uses a connection path between a user and an item in the knowledge graph to extract a rule which can represent the user interest preference and the item attribute characteristics most in a specified recommendation scene. The Bi-LSTM model is then employed to learn the semantic features of the paths generated by the rules, and an attention mechanism is employed to distinguish between the user's preferences for different rules and those that best represent the specified items. Then, a local information sensing module is designed to learn the close-range neighborhood characteristics of the user and the project; and finally, integrating the user and item expressions learned by the global information perception module and the local information perception module through an aggregation algorithm, and implementing prediction. The experimental results on three real data sets prove that the performance of the method is obviously improved compared with that of a benchmark method.
Fig. 8 is an architecture diagram of a knowledge-graph-based multi-context-aware recommendation system according to an embodiment of the present invention, as shown in fig. 8, including:
a rule-based path extraction module 810 for determining a path between the user and each of the plurality of items according to the interaction records generated between the user and the plurality of items; extracting rules for generating paths between the users and the items from the paths according to the association among the plurality of items; the items are entities in a knowledge graph; the path is used for explaining how the user interacts with the item, and the rule is used for reflecting the interest preference of the user on the entity; filtering the rules based on chi-square distribution according to the number of paths which are connected correctly and incorrectly by the rules and the number of paths which are not connected correctly and incorrectly by the rules, and keeping the high-quality rules; determining all paths between the user and the project based on the high-quality rule, and eliminating noise paths in all the paths based on the weight of each path to obtain path samples between the user and the project;
a global environment sensing module 820, configured to learn low-dimensional embedded representations of paths in the path sample by using a bidirectional long and short term memory network, and aggregate the low-dimensional embedded representations of the paths corresponding to the rules to obtain long-distance multivariate semantic connection information between users and projects based on a global environment;
a local environment sensing module 830, configured to learn different attention scores of a plurality of neighborhood entities of the user and a plurality of neighborhood entities of the project based on the graph attention network, so as to aggregate corresponding neighborhood entities by the different attention scores, so as to obtain close-range neighborhood information of the user and the project based on the local environment, respectively;
the scoring prediction module 840 is used for aggregating long-distance multi-element semantic connection information between users and projects, close-range neighborhood information of the users and close-range neighborhood information of the projects to obtain comprehensive characterization vectors of the users and the projects; and inputting the comprehensive characterization vectors of the users and the items into a multilayer perceptron to obtain a prediction score between the users and the items, and recommending the corresponding items for the users based on the prediction score.
It can be understood that detailed functional implementation of each module in fig. 8 can refer to the description in the foregoing method embodiment, and is not described herein again.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (10)
1. A knowledge graph-based multi-element environment perception recommendation method is characterized by comprising the following steps:
determining a path between the user and each project according to the interaction records generated between the user and the projects; extracting rules for generating paths between the users and the items from the paths according to the association among the plurality of items; the items are entities in a knowledge graph; the path is used for explaining how the user interacts with the item, and the rule is used for reflecting the interest preference of the user on the entity;
filtering the rules based on chi-square distribution according to the number of paths which are connected correctly and incorrectly by the rules and the number of paths which are not connected correctly and incorrectly by the rules, and keeping the high-quality rules;
determining all paths between the user and the project based on the high-quality rule, and eliminating noise paths in all the paths based on the weight of each path to obtain path samples between the user and the project;
learning the low-dimensional embedded representation of each path in the path sample by using a bidirectional long and short term memory network, and aggregating the low-dimensional embedded representation of the path corresponding to each rule to obtain long-distance multi-semantic connection information between the user and the project based on the global environment;
learning different attention scores of a plurality of neighborhood entities of the user and a plurality of neighborhood entities of the project based on a graph attention network, and aggregating the corresponding neighborhood entities through the different attention scores to respectively acquire close neighborhood information of the user and the project based on the local environment;
aggregating long-distance multi-element semantic connection information between users and projects, close-range neighborhood information of the users and close-range neighborhood information of the projects to obtain comprehensive characterization vectors of the users and the projects;
and inputting the comprehensive characterization vectors of the users and the items into a multilayer perceptron to obtain a prediction score between the users and the items, and recommending the corresponding items for the users based on the prediction score.
2. The method of claim 1, wherein according to the interaction records generated between the user and the plurality of items, a bidirectional path search strategy is adopted to perform breadth-first search strategies from the user and the items respectively to determine all connection paths between the user and the items, and the connection relationship of each path is taken as a rule;
given a rule, orderRepresenting all user-item pairs connected by interactive relations in the knowledge graph, u representing a user, and v representing an item;representing all user-item pairs in the knowledge-graph connected by the path generated by the rule; then:
tnnum=m*n-tpnum-fpnum-fnnumrepresenting the number of erroneous user-item pairs that are not connected by a rule; m represents the total number of users, and n represents the total number of items;
the obtained tpnum、fpnum、fnnum、tnnumForm a List table [ (tp)num,fpnum),(fnnum,tnnum)]Further, a chi-square detection method is adopted to carry out statistics on the list table to obtain a statistic value; if the statistical value is smaller than a set threshold value, the quality of the rule is high, and the rule is reserved; otherwise, the rule is removed.
3. The method of claim 1, wherein the embedded feature representation of each node in the knowledge-graph is learned using a factorization approach; calculating the similarity between two adjacent nodes in the path generated by each rule by adopting cosine similarity based on the embedded characteristic representation of each node; averaging the similarity scores between two adjacent nodes in each path, and taking the average as the priority score of the path; finally, for each rule, the k paths with the highest priority are selected to represent the path samples between the user and the item with respect to the rule.
4. A method according to any one of claims 1 to 3, characterized in that the embedded representation of a given path of length s is p ═ eu,ew1,ew2,...,evUsing forward hidden state sequence of each entity on the Bi-LSTM learning pathAnd reverse hidden state sequenceThe mutual information of the learning node and the forward neighbor and the backward neighbor is used; e.g. of the typeuAnd evRepresenting users and items on the path, respectively, ew1、ew2.., other entities on the path;
connecting the forward and reverse state outputs of any unit in the bidirectional long-short term memory network (Bi-LSTM) as an entity e in the pathjFinal output h after Bi-LSTM treatmentj(ii) a Wherein e isj∈eu、ew1、ew2、...、ev;
For a path p with the length s, the output of the path p after the Bi-LSTM is a matrix [ h ] formed by the output states of all entitiesu,hw1,hw2,...,hv](ii) a Merging the matrixes by adopting a pooling method to obtain a final path embedded representation ep(ii) a Embedding paths into representation epOutput h processed by Bi-LSTM respectively with user and project nodeu、hvConnecting to obtain the representation e of the user and the project on the pathupAnd evp;
Aggregating all path-based user and item representations contained in each rule through a pooling method to obtain rule-based user and item embedded vector representation eurAnd evr;
Order toRepresenting rule-based user-embedded vector representation [ e ]ur1,eur2,...,eurl]A two-dimensional matrix is formed, wherein l is the number of set rules in the recommended scene, eurjFor rule j-based users embedding vector representations, d is the length of each rule vector representation;
the regular attention coefficient score is obtained by the following formulaur:scoreur=LeakyReLU(W1Eur) (ii) a Wherein,is a first linear transformation weight matrix; LeakyReLU is an activation function;
normalizing the attention coefficients of all the rules by utilizing a softmax function to obtain a final attention weight vector of each rule: attenur=softmax(W2scoreur) (ii) a Wherein, representing a second linear transformation weight matrix; attenurRepresenting an attention weight vector;
finally, user u represents e based on rule embeddingur' is: e.g. of the typeur′=attenurEur;
Likewise, a rule-based embedded representation e of the item v can be obtainedvr′。
5. The method of claim 4, wherein the instructions are executedRepresenting a local neighborhood of user u;
for user u, user u and its neighborhood are calculated respectivelySimilarity coefficient between:in the formula, W3Is the weight matrix of the third linear transformation, W4Is a weight matrix of a fourth linear transformation [ | | · [ ] | ]]Representation for user u and item viSplicing the linearly transformed features, and mapping the spliced high-dimensional features to a real number by using an a (-) function;
normalizing the attention coefficient by adopting a softmax function to obtain the normalized attention coefficient
Weighting and summing the characteristics of the neighborhood according to the normalized attention coefficient to obtain the characterization e of the user u aggregated with the local neighborhood characteristicsut:W5Is a weight matrix of a fifth linear transformation;
6. The method of claim 5, wherein the long-distance multi-semantic connection information e between the user and the project is obtained by means of connection aggregationur' and evr', user's close proximity information eutAnd close proximity information e of the itemvtPolymerizing to obtain a connected polymerization vector aggconcat:
aggconcat=σ(W8(eur′||eut||evr′||evt)+b)
Agg is preparedconcatAs a comprehensive characterization vector e for users and itemsuv;
E is to beuvInputting the data into a multi-layer perceptron to obtain the prediction scores of the users and the projects
In the formula, W8、W9A weight matrix representing the eighth linear transformation and a weight matrix representing the ninth linear transformation, and b represents a bias of the linear transformation.
7. A knowledge-graph-based multivariate environment-aware recommendation system, comprising:
the rule-based path extraction module is used for determining a path between a user and each project according to an interaction record generated between the user and a plurality of projects; extracting rules for generating paths between the users and the items from the paths according to the association among the plurality of items; the items are entities in a knowledge graph; the path is used for explaining how the user interacts with the item, and the rule is used for reflecting the interest preference of the user on the entity; filtering the rules based on chi-square distribution according to the number of paths which are connected correctly and incorrectly by the rules and the number of paths which are not connected correctly and incorrectly by the rules, and keeping the high-quality rules; determining all paths between the user and the project based on the high-quality rule, and eliminating noise paths in all the paths based on the weight of each path to obtain path samples between the user and the project;
the global environment perception module is used for learning the low-dimensional embedded representation of each path in the path sample by utilizing a bidirectional long and short term memory network and aggregating the low-dimensional embedded representation of the path corresponding to each rule to obtain long-distance multi-element semantic connection information between the user and the project based on the global environment;
the local environment sensing module is used for learning different attention scores of a plurality of neighborhood entities of the user and a plurality of neighborhood entities of the project based on the graph attention network so as to aggregate the corresponding neighborhood entities through the different attention scores to respectively acquire close-range neighborhood information of the user and the project based on the local environment;
the scoring prediction module is used for aggregating long-distance multi-element semantic connection information between the users and the projects, close-range neighborhood information of the users and close-range neighborhood information of the projects to obtain comprehensive characterization vectors of the users and the projects; and inputting the comprehensive characterization vectors of the users and the items into a multilayer perceptron to obtain a prediction score between the users and the items, and recommending the corresponding items for the users based on the prediction score.
8. The system of claim 7, wherein the rule-based path extraction moduleThe method comprises the following steps that a block is used for determining all connection paths between a user and a plurality of items by adopting a bidirectional path searching strategy and executing a breadth-first searching strategy from the user and the items respectively according to interaction records generated between the user and the items, and the connection relation of each path is used as a rule; given a rule, orderRepresenting all user-item pairs connected by interactive relations in the knowledge graph, u representing a user, and v representing an item;representing all user-item pairs in the knowledge-graph connected by the path generated by the rule; then:indicating the number of correct user-item pairs that are connected by the rule;indicating the number of correct user-item pairs that are not connected by the rule;representing the number of erroneous user-item pairs that are connected by the rule; tnnum=m*n-tpnum-fpnum-fnnumRepresenting the number of erroneous user-item pairs that are not connected by a rule; m represents the total number of users, and n represents the total number of items; the obtained tpnum、fpnum、fnnum、tnnumForm a series table [ (tp)num,fpnum),(fnnum,tnnum)]Further, a chi-square detection method is adopted to carry out statistics on the list table to obtain a statistic value; if the statistical value is smaller than a set threshold value, the quality of the rule is high, and the rule is reserved; otherwise, the rule is removed.
9. The system according to claim 7 or 8, wherein the global context awareness module obtains long-distance multivariate semantic connection information between the user and the project based on the global context, and specifically comprises:
given an embedded representation of a path of length s, p ═ eu,ew1,ew2,...,evUsing forward hidden state sequence of each entity on the Bi-LSTM learning pathAnd reverse hidden state sequenceThe mutual information of the learning node and the forward neighbor and the backward neighbor is used; e.g. of the typeuAnd evRepresenting users and items on the path, respectively, ew1、ew2.., other entities on the path;
connecting the forward and reverse state outputs of any unit in the bidirectional long-short term memory network (Bi-LSTM) as an entity e in the pathjFinal output h after Bi-LSTM treatmentj(ii) a Wherein e isj∈eu、ew1、ew2、...、ev;
For a path p with the length s, the output of the path p after the Bi-LSTM is a matrix [ h ] formed by the output states of all entitiesu,hw1,hw2,...,hv](ii) a Merging the matrixes by adopting a pooling method to obtain a final path embedded representation ep(ii) a Embedding paths into representation epOutput h processed by Bi-LSTM respectively with user and project nodeu、hvConnecting to obtain the representation e of the user and the project on the pathupAnd evp;
Aggregating all path-based user and item representations contained in each rule through a pooling method to obtain rule-based user and item embedded vector representation eurAnd evr;
Order toRepresenting rule-based user-embedded vector representation [ e ]ur1,eur2,...,eurl]A two-dimensional matrix is formed, wherein l is the number of set rules in the recommended scene, eurjFor rule j-based users embedding vector representations, d is the length of each rule vector representation;
the regular attention coefficient score is obtained by the following formulaur:scoreur=LeakyReLU(W1Eur) (ii) a Wherein,is a first linear transformation weight matrix; LeakyReLU is an activation function;
normalizing the attention coefficients of all the rules by utilizing a softmax function to obtain a final attention weight vector of each rule: attenur=softmax(W2scoreur) (ii) a Wherein, representing a second linear transformation weight matrix; attenurRepresenting an attention weight vector;
finally, user u represents e based on rule embeddingur' is: e.g. of a cylinderur′=attenurEur;
Likewise, a rule-based embedded representation e of the item v can be obtainedvr′。
10. The system according to claim 7 or 8, wherein the local environment awareness module obtains close proximity information of users and items based on local environment, and specifically comprises:
for user u, user u and its neighborhood are calculated respectivelySimilarity coefficient between:in the formula, W3Is the weight matrix of the third linear transformation, W4Is a weight matrix of a fourth linear transformation [ | | · [ ] | ]]Representation for user u and item viSplicing the linearly transformed features, and mapping the spliced high-dimensional features to a real number by using an a (-) function;
normalizing the attention coefficient by adopting a softmax function to obtain the normalized attention coefficient
Weighting and summing the characteristics of the neighborhood according to the normalized attention coefficient to obtain the characterization e of the user u aggregated with the local neighborhood characteristicsut:W5Is a weight matrix of a fifth linear transformation;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210079690.9A CN114461907B (en) | 2022-01-24 | 2022-01-24 | Knowledge graph-based multi-element environment perception recommendation method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210079690.9A CN114461907B (en) | 2022-01-24 | 2022-01-24 | Knowledge graph-based multi-element environment perception recommendation method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114461907A true CN114461907A (en) | 2022-05-10 |
CN114461907B CN114461907B (en) | 2024-06-21 |
Family
ID=81411534
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210079690.9A Active CN114461907B (en) | 2022-01-24 | 2022-01-24 | Knowledge graph-based multi-element environment perception recommendation method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114461907B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115878904A (en) * | 2023-02-22 | 2023-03-31 | 深圳昊通技术有限公司 | Intellectual property personalized recommendation method, system and medium based on deep learning |
CN116842199A (en) * | 2023-09-01 | 2023-10-03 | 东南大学 | Knowledge graph completion method based on multi-granularity hierarchy and dynamic embedding |
CN117610662A (en) * | 2024-01-19 | 2024-02-27 | 江苏天人工业互联网研究院有限公司 | Knowledge graph embedding method for extracting representative sub-graph information through GAT |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190236469A1 (en) * | 2018-02-01 | 2019-08-01 | International Business Machines Corporation | Establishing a logical connection between an indirect utterance and a transaction |
CN110516160A (en) * | 2019-08-30 | 2019-11-29 | 中国科学院自动化研究所 | User modeling method, the sequence of recommendation method of knowledge based map |
US20210248461A1 (en) * | 2020-02-11 | 2021-08-12 | Nec Laboratories America, Inc. | Graph enhanced attention network for explainable poi recommendation |
CN113378047A (en) * | 2021-06-10 | 2021-09-10 | 武汉大学 | Multi-aspect enhancement-based graph neural network recommendation method |
-
2022
- 2022-01-24 CN CN202210079690.9A patent/CN114461907B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190236469A1 (en) * | 2018-02-01 | 2019-08-01 | International Business Machines Corporation | Establishing a logical connection between an indirect utterance and a transaction |
CN110516160A (en) * | 2019-08-30 | 2019-11-29 | 中国科学院自动化研究所 | User modeling method, the sequence of recommendation method of knowledge based map |
US20210248461A1 (en) * | 2020-02-11 | 2021-08-12 | Nec Laboratories America, Inc. | Graph enhanced attention network for explainable poi recommendation |
CN113378047A (en) * | 2021-06-10 | 2021-09-10 | 武汉大学 | Multi-aspect enhancement-based graph neural network recommendation method |
Non-Patent Citations (1)
Title |
---|
张素琪;许馨匀;佘士耀;任珂可;: "基于双用户视角与知识图谱注意力网络的推荐模型", 现代计算机, no. 13, 5 May 2020 (2020-05-05) * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115878904A (en) * | 2023-02-22 | 2023-03-31 | 深圳昊通技术有限公司 | Intellectual property personalized recommendation method, system and medium based on deep learning |
CN115878904B (en) * | 2023-02-22 | 2023-06-02 | 深圳昊通技术有限公司 | Intellectual property personalized recommendation method, system and medium based on deep learning |
CN116842199A (en) * | 2023-09-01 | 2023-10-03 | 东南大学 | Knowledge graph completion method based on multi-granularity hierarchy and dynamic embedding |
CN116842199B (en) * | 2023-09-01 | 2023-12-26 | 东南大学 | Knowledge graph completion method based on multi-granularity hierarchy and dynamic embedding |
CN117610662A (en) * | 2024-01-19 | 2024-02-27 | 江苏天人工业互联网研究院有限公司 | Knowledge graph embedding method for extracting representative sub-graph information through GAT |
Also Published As
Publication number | Publication date |
---|---|
CN114461907B (en) | 2024-06-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wei et al. | MMGCN: Multi-modal graph convolution network for personalized recommendation of micro-video | |
CN111931062B (en) | Training method and related device of information recommendation model | |
CN114461907B (en) | Knowledge graph-based multi-element environment perception recommendation method and system | |
Tong et al. | Leveraging domain context for question answering over knowledge graph | |
Zhang et al. | Cross-domain recommendation with semantic correlation in tagging systems | |
Wang et al. | Group-aware long-and short-term graph representation learning for sequential group recommendation | |
CN113918834B (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
CN113918832A (en) | Graph convolution collaborative filtering recommendation system based on social relationship | |
CN113918833A (en) | Product recommendation method realized through graph convolution collaborative filtering of social network relationship | |
CN116401380B (en) | Heterogeneous knowledge graph-oriented contrast learning prediction method and system | |
CN116431914A (en) | Cross-domain recommendation method and system based on personalized preference transfer model | |
Zhiyuli et al. | Bookgpt: A general framework for book recommendation empowered by large language model | |
Patro et al. | Knowledge-based preference learning model for recommender system using adaptive neuro-fuzzy inference system | |
Sakurai et al. | Deep reinforcement learning-based music recommendation with knowledge graph using acoustic features | |
Qiao et al. | Bi-channel Multiple Sparse Graph Attention Networks for Session-based Recommendation | |
Deng et al. | A Trust-aware Neural Collaborative Filtering for Elearning Recommendation. | |
Cardoso et al. | Vulcont: A recommender system based on context history ontology | |
Banouar et al. | Enriching SPARQL queries by user preferences for results adaptation | |
Qi et al. | ScratchGAN: Network representation learning for scratch with preference‐based generative adversarial nets | |
Shu et al. | Self-Supervised Hypergraph Learning for Enhanced Multimodal Representation | |
Mecheri et al. | Deep learning based web service recommendation methods: A survey | |
Wu et al. | A Review of User Profiling Based on Social Networks | |
Wang et al. | BERT-based aggregative group representation for group recommendation | |
Su et al. | FPLV: Enhancing recommender systems with fuzzy preference, vector similarity, and user community for rating prediction | |
Wang et al. | Task Recommendation via Heterogeneous Multi-modal Features and Decision Fusion in Mobile Crowdsensing |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |