CN114820139B - Multi-user recommendation system based on knowledge graph path reasoning - Google Patents
Multi-user recommendation system based on knowledge graph path reasoning Download PDFInfo
- Publication number
- CN114820139B CN114820139B CN202210583484.1A CN202210583484A CN114820139B CN 114820139 B CN114820139 B CN 114820139B CN 202210583484 A CN202210583484 A CN 202210583484A CN 114820139 B CN114820139 B CN 114820139B
- Authority
- CN
- China
- Prior art keywords
- path
- user
- knowledge graph
- module
- item
- 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.)
- Active
Links
- 230000003993 interaction Effects 0.000 claims abstract description 39
- 238000011156 evaluation Methods 0.000 claims abstract description 15
- 238000010276 construction Methods 0.000 claims abstract description 11
- 239000013598 vector Substances 0.000 claims description 38
- 238000003062 neural network model Methods 0.000 claims description 16
- 238000000034 method Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 11
- 125000004122 cyclic group Chemical group 0.000 claims description 9
- 230000002452 interceptive effect Effects 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 5
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 4
- 238000005065 mining Methods 0.000 claims description 4
- 230000000306 recurrent effect Effects 0.000 claims 2
- 230000002776 aggregation Effects 0.000 abstract description 4
- 238000004220 aggregation Methods 0.000 abstract description 4
- 230000007246 mechanism Effects 0.000 abstract description 4
- 238000011176 pooling Methods 0.000 abstract description 4
- 230000006870 function Effects 0.000 description 12
- 230000004075 alteration Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 230000015654 memory Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- 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
- G06N3/048—Activation functions
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Development Economics (AREA)
- Strategic Management (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Economics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Entrepreneurship & Innovation (AREA)
- Databases & Information Systems (AREA)
- Animal Behavior & Ethology (AREA)
- Game Theory and Decision Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a multi-user recommendation system based on knowledge graph path reasoning, which comprises a knowledge graph construction module, a path reasoning module and a scoring prediction module; the knowledge graph construction module acquires user interaction historical data, constructs a knowledge graph G and transmits the knowledge graph G to the path reasoning module; the path reasoning module generates a plurality of relation paths between the user and the target item according to the knowledge graph G and transmits the relation paths to the scoring prediction module; and the evaluation prediction module is used for performing evaluation prediction on the received relation paths and outputting an item recommendation list to a user according to the evaluation prediction result. The invention digs more abundant user interest hidden semantic information by constructing the knowledge graph of the user and the item, predicts the potential common preference of a plurality of users by utilizing the combination reasoning of a plurality of paths on the knowledge graph, and gives a recommendation list maximizing interest preference and diversification while distinguishing the preference degree of different users by utilizing pooling aggregation operation with an attention mechanism among different paths.
Description
Technical Field
The invention relates to the field of user recommendation, in particular to a multi-user recommendation system based on knowledge graph path reasoning.
Background
Most of the recommendation technologies studied in academia and industry are personalized recommendation systems about single users, however, the recommended objects are often groups of more than one person, so that group recommendation technologies are generated, the group recommendation systems form groups by aggregating groups with similar interests, and then group aggregation is realized by fusing preferences of all members in the groups, for example, users who frequently browse or purchase similar products are aggregated into one group by mining user interaction history. However, in real life, there are naturally occurring groups like families, classes, companies, etc., and interest correlation between these groups may be small, and in some occasions, these groups may only purchase resources through the same device ID or the same login account. For these multiple users, conventional group recommendation systems are not applicable, for which reason multiple user recommendation systems are derived.
Typical application scenarios for multi-user recommendation systems are: 1) IPTV field: a plurality of family members of the household television can only watch television on line through the same set top box IP; 2) Video mobile client: a plurality of clients share the same member account to log in the love art and make a message to select the video resources of interest; 3) E-commerce platform: in order to enjoy discount service, family members share the same member account number to log in an e-commerce platform to purchase goods online, and the like. For the users, the interest difference among the internal members is large, however, a multi-user group is constructed due to natural factors, and commodities are selected and purchased through the same account number or the same equipment IP. Recommending content for these groups that meets the interests of members within the group is known as a multi-user recommendation system.
How to construct a multi-user recommendation system, so that the information overload massive resources can be found through the same equipment IP or the same account, thereby meeting the requirements of a plurality of users and reflecting personalized recommendation, and the following problems need to be solved:
1) Maximizing common interest preferences among members;
2) Requiring diversification of recommendation results to meet personalized requirements among members;
3) The lack of social relations among groups, how to utilize the interaction history of the existing users and the items to mine more effective information;
Disclosure of Invention
The invention aims to provide a multi-user recommendation system based on knowledge graph path reasoning, which comprises a knowledge graph construction module, a path reasoning module and a scoring prediction module;
The knowledge graph construction module acquires user interaction historical data, constructs a knowledge graph G and transmits the knowledge graph G to the path reasoning module;
The step of constructing the knowledge graph G includes:
a) Constructing a project association directed graph G 1; the entity of the item association directed graph G 1 comprises interaction items and candidate items, and the edge set comprises an association relationship between the entities;
project association directed graph G 1＝{(h,r,t)|h,t∈I1,r∈R1, obtained by modeling association data between candidate projects and interacted projects;
Wherein the tuple (h, r, t) indicates that there is a relationship r between the head node h and the tail node t; i 1 is a set of all candidate items; r 1 is a relation set b) to construct a directed graph G 2 of user interaction with the project; the entities of the interactive directed graph G 2 include users and items; in the interaction directed graph G 2, if the interaction exists between the user and the item, an edge exists between the user and the item;
the directed graph G 2＝{(u,act,i)|u∈U,i∈I2,act∈R2 of the interaction between the user and the project is obtained by modeling the interaction record between the user and the project;
Wherein the tuple (u, act, i) represents that an interactive relationship act exists between the user u and the item i; And/> Respectively representing the user and the item set, M and N respectively representing the number of users and the number of items,/>R 2 represents a set of interaction relationships.
C) And fusing the project association directed graph G 1 and the interaction directed graph G 2 of the user and the project to obtain a knowledge graph G.
The knowledge graph g= { (h, R, t) |h, t E, R E R }, wherein set e=u U I 1, set r=r 1∪R2. The path reasoning module generates a plurality of relation paths between the user and the target item according to the knowledge graph G and transmits the relation paths to the scoring prediction module;
The step of generating a relationship path between the plurality of users and the target item by the path reasoning module comprises the following steps:
1) Recording paths between users and items on a knowledge graph G as a sequence of entities and relations Where e 1＝u,eL＝p;(el,rl,el+1) is the first tuple on path p, L-1 represents the number of all tuples on path p;
2) For a given relationship path p k on the knowledge graph G, the path inference module segments each entity in the relationship path p k into a corresponding type and value, initializes the type and value into two vectors, and records the two vectors as D represents the dimension size; the initial value of k is 1;
3) Embedding a vector r l with the same dimension in a path p k, updating a relation path p 'k to be [ e 1,r1,e2,...,el-1,rl-1 ], and enabling each element in the relation path p' k to represent an entity or a class of relation;
4) Mining sequence information in the relationship path p' k by using the cyclic neural network model stored on the path inference module, thereby generating an encoded path p k;
the circulating neural network model is an LSTM neural network model and comprises an input layer, a hiding layer and an output layer; the hidden layer comprises an input door, a forgetting door and an output door;
the input of the cyclic neural network model comprises entity types, entity values and relation vectors, and the output is an encoded path.
The input vector x l-1 and the output vector h L of the cyclic neural network model respectively meet the following formulas:
zl＝tanh(Wzxl+Whhl-1+bz) (2)
fl＝σ(Wfxl+Whhl-1+bf) (3)
il＝σ(Wixl+Whhl-1+bi) (4)
ol＝σ(Woxl+Whhl-1+bo) (5)
cl＝fl⊙cl-1+il⊙zl (6)
hl＝ol⊙tanh(cl) (7)
Wherein c l∈Rd,zl∈Rd represents a unit storage state vector and an information conversion module, and d' represents the number of hidden units; i l、ol and f l represent input, output and forget gates, respectively, and W z、Wi、Wf、Wo∈Rd'×3d,Wh∈Rd'×d' represents a mapping coefficient matrix; b z、bi、bf、bo is the bias vector; sigma (-) is an activation function, by which is meant the element-level product between the two vectors; e l-1、e′l-1、rl-1 represents entity type, entity value, and relationship vector, respectively.
5) Let k=k+1 and return to step 2) until the encoded path set P (u, i) = { P 1,p2,...,pK }, is obtained.
And the evaluation prediction module is used for performing evaluation prediction on the received relation paths and outputting an item recommendation list to a user according to the evaluation prediction result.
The items in the item recommendation list are predicted according to scoresThe Top N items in the rank-order are decremented (Top-N recommendation).
The step of evaluating and predicting the received relation paths by the evaluation prediction module comprises the following steps:
I) Establishing a prediction function, namely:
In the method, in the process of the invention, To evaluate the prediction, f θ represents a model with a parameter θ, which indicates whether user u interacted with item i was inferred from connectivity ρ (u, i);
II) calculating state predicted values, namely:
Wherein W 1,W2 represents the coefficient weights of the 1 st full-connection layer and the 2 nd full-connection layer respectively; relu denotes an activation function; τ= (u, act, i) represents a tuple; s (τ|p k)＝sk, representing the predictive score of the kth path from node u to node i).
III) projecting the state predicted value into a prediction function to obtain an evaluation predicted resultNamely:
Where σ ()' is the activation function;
Wherein, the parameter g (s 1,s2,...,sK) is as follows:
Wherein γ is a hyper-parameter controlling each exponential weight; { s 1,s2,...,sK } represents the respective predictive scores of the total K paths from node u to node i. s k represents the prediction score of the kth path from node u to node i.
The invention provides a multi-user recommendation system based on knowledge graph path reasoning, which excavates more abundant user interest hidden semantic information by constructing a knowledge graph of users and items, predicts potential common preferences of a plurality of users by utilizing multiple path combination reasoning on the knowledge graph, and gives a recommendation list maximizing interest preferences and diversification while distinguishing different user preference degrees by utilizing pooling aggregation operation with a attention mechanism among different paths.
According to the method, a Knowledge Graph (KG) is established according to the historical interaction data of the user and the association information between the items, reasonable paths between the user and the items are extracted from the KG, and the sequential dependency relationship between entities such as the user items and the complex relationship between paths connecting the user and the items are modeled. Simulating the sequential dependency relationship of entities and relationships by adopting a Long and Short Term Memory (LSTM) network;
The invention designs the aggregation operation of the weighted pool, distinguishes different contribution values of a plurality of paths between user-project interaction tuples, and has a certain attention mechanism for distinguishing interest preference among different users; compared with the traditional operation of taking the average value of all paths, the recommendation algorithm realizes the path-level interpretability;
According to the invention, the knowledge graph is applied to the multi-user recommendation method for the first time, the traditional multi-user recommendation model is expanded, a knowledge-aware path circulation neural network is provided, more candidate items meeting the interests of the user are inferred through the history record of the user, the long-tail items are conveniently mined, and the diversity of recommendation results is improved;
Besides learning the relation expression of the user and the project, the invention also extracts complex and various association relations among project products. Meanwhile, the diversity of paths is guaranteed, the setting of parameters of the paths Chi Huachao is in direct proportion to the score of each path in the back propagation process, and more flexibility is provided for final prediction.
Drawings
Fig. 1 is a general flow chart of the present solution.
Fig. 2 is a schematic diagram of the structural components of the model according to the present embodiment.
Fig. 3 is a schematic diagram of a core knowledge graph path reasoning model.
Detailed Description
The present invention is further described below with reference to examples, but it should not be construed that the scope of the above subject matter of the present invention is limited to the following examples. Various substitutions and alterations are made according to the ordinary skill and familiar means of the art without departing from the technical spirit of the invention, and all such substitutions and alterations are intended to be included in the scope of the invention.
Example 1:
Referring to fig. 1 to 3, a multi-user recommendation system based on knowledge-graph path reasoning includes a knowledge-graph construction module, a path reasoning module and a scoring prediction module;
The knowledge graph construction module acquires user interaction historical data, constructs a knowledge graph G and transmits the knowledge graph G to the path reasoning module;
The step of constructing the knowledge graph G includes:
a) Constructing a project association directed graph G 1; the entity of the item association directed graph G 1 comprises interaction items and candidate items, and the edge set comprises an association relationship between the entities;
project association directed graph G 1＝{(h,r,t)|h,t∈I1,r∈R1, obtained by modeling association data between candidate projects and interacted projects;
Wherein the tuple (h, r, t) indicates that there is a relationship r between the head node h and the tail node t; i 1 is a set of all candidate items; r 1 is a set of relationships.
B) Constructing a user interaction directed graph G 2 with the project; the entities of the interactive directed graph G 2 include users and items; in the interaction directed graph G 2, if the interaction exists between the user and the item, an edge exists between the user and the item;
the directed graph G 2＝{(u,act,i)|u∈U,i∈I2,act∈R2 of the interaction between the user and the project is obtained by modeling the interaction record between the user and the project;
Wherein the tuple (u, act, i) represents that an interactive relationship act exists between the user u and the item i; And Respectively representing the user and the item set, M and N respectively representing the number of users and the number of items,/>R 2 represents a set of interaction relationships.
C) And fusing the project association directed graph G 1 and the interaction directed graph G 2 of the user and the project to obtain a knowledge graph G.
The knowledge graph g= { (h, R, t) |h, t E, R E R }, wherein set e=u U I 1, set r=r 1∪R2. .
The path reasoning module generates a plurality of relation paths between the user and the target item according to the knowledge graph G and transmits the relation paths to the scoring prediction module;
The step of generating a relationship path between the plurality of users and the target item by the path reasoning module comprises the following steps:
1) Recording paths between users and items on a knowledge graph G as a sequence of entities and relations Where e 1＝u,eL＝p;(el,rl,el+1) is the first tuple on path p, L-1 represents the number of all tuples on path p;
2) For a given relationship path p k on the knowledge graph G, the path inference module segments each entity in the relationship path p k into a corresponding type and value, initializes the type and value into two vectors, and records the two vectors as D represents the dimension size; the initial value of k is 1;
3) Embedding a vector r l with the same dimension in a path p k, updating a relation path p 'k to be [ e 1,r1,e2,...,el-1,rl-1 ], and enabling each element in the relation path p' k to represent an entity or a class of relation;
4) Mining sequence information in the relationship path p' k by using the cyclic neural network model stored on the path inference module, thereby generating an encoded path p k;
the circulating neural network model is an LSTM neural network model and comprises an input layer, a hiding layer and an output layer; the hidden layer comprises an input door, a forgetting door and an output door;
the input of the cyclic neural network model comprises entity types, entity values and relation vectors, and the output is an encoded path.
The input vector x l-1 and the output vector h L of the cyclic neural network model respectively meet the following formulas:
zl＝tanh(Wzxl+Whhl-1+bz) (2)
fl＝σ(Wfxl+Whhl-1+bf) (3)
il＝σ(Wixl+Whhl-1+bi) (4)
ol＝σ(Woxl+Whhl-1+bo) (5)
cl＝fl⊙cl-1+il⊙zl (6)
hl＝ol⊙tanh(cl) (7)
Wherein c l∈Rd,zl∈Rd represents a unit storage state vector and an information conversion module, and d' represents the number of hidden units; i l、ol and f l represent input, output and forget gates, respectively, and W z、Wi、Wf、Wo∈Rd'×3d,Wh∈Rd'×d' represents a mapping coefficient matrix; b z、bi、bf、bo is the bias vector; sigma (-) is an activation function, by which is meant the element-level product between the two vectors; e l-1、e′l-1、rl-1 represents entity type, entity value, and relationship vector, respectively.
5) Let k=k+1 and return to step 2) until the encoded path set P (u, i) = { P 1,p2,...,pK }, is obtained.
And the evaluation prediction module is used for performing evaluation prediction on the received relation paths and outputting an item recommendation list to a user according to the evaluation prediction result.
The items in the item recommendation list are predicted according to the scoresThe Top N items in the rank-order are decremented (Top-N recommendation).
The step of evaluating and predicting the received relation paths by the evaluation prediction module comprises the following steps:
I) Establishing a prediction function, namely:
In the method, in the process of the invention, To evaluate the predicted outcome; f θ denotes a model with a parameter θ, which indicates whether user u interacted with item i is inferred from connectivity ρ (u, i).
II) calculating state predicted values, namely:
Wherein W 1,W2 represents the coefficient weights of the 1 st full-connection layer and the 2 nd full-connection layer respectively; relu denotes an activation function; where τ= (u, act, i) represents a tuple; s (τ|p k)＝sk, representing the predictive score of the kth path from node u to node i;
III) projecting the state predicted value into a prediction function to obtain an evaluation predicted result Namely:
Where σ ()' is the activation function;
Wherein, the parameter g (s 1,s2,...,sK) is as follows:
Wherein γ is a hyper-parameter controlling each exponential weight; { s 1,s2,...,sK } represents the respective predictive scores of the total K paths from node u to node i. s k represents the prediction score of the kth path from node u to node i.
Example 2:
a multi-user recommendation system based on knowledge graph path reasoning comprises a knowledge graph construction module, a path reasoning module and a scoring prediction module.
Knowledge graph construction module:
the knowledge graph construction module firstly constructs a project association directed graph G 1, wherein entities are interaction items and candidate items, and an edge set comprises association relations (such as a same type relation and a collocation relation) between the entities; and secondly, constructing a directed graph G 2 of the interaction between the user and the project, wherein the entity is the user and the project, and if the interaction exists between the user and the project, edges exist. And fusing the G 1 and the G 2 to obtain a unified knowledge graph G.
Specifically, the association data between candidate items and interacted items is first modeled as a directed graph G 1＝{(h,r,t)|h,t∈I1,r∈R1, each tuple (h, r, t) indicating that there is a relationship r between a head node h and a tail node t, due to the presence of different products: the association of similar types, alternatives, combinations, etc. with different strengths needs to be specific to a particular scene. For example, in an IPTV home online television scenario, for a movie asset, a episode a is similar to episode B, and episode B and episode C are associated (the main angles of B and C are the same actor).
Next, the user interaction record with the item is modeled as a bipartite graph G 2＝{(u,act,i)|u∈U,i∈I2,act∈R2, usingAnd/>Representing the user, the set of items, respectively, where M and N represent the number of users and the number of items, respectively, and the tuple τ= (u, act, i) is used to represent the interaction relationship between the user and the items, for example: purchasing, collecting, commenting, ordering, etc.
Next, the user interaction graph and the project relation graph are fused to obtain a unified knowledge graph g= { (h, R, t) |h, t E, R E R }, wherein the set e=u ∈i 1 and the set r=r 1∪R2.
And a path reasoning module based on the knowledge graph:
the path reasoning module defines paths between user arrival items on G as a sequence of entities and relationships: Where e 1＝u,eL＝p;(el,rl,el+1) is the first tuple on path p, L-1 represents the number of all tuples on path p.
The path inference module segments each entity in p k into a corresponding type and value (e.g., entity type is user, value is min) and initializes two vectors for a given path p k D represents the dimension size. The paths in the homonymy path p k are embedded as vectors r l of the same dimension, thus resulting in the characterization of the relationship path p k as: [ e 1,r1,e2,...,el-1,rl-1 ] each element represents an entity or class of relationships.
After a description path of an embedded sequence is obtained, sequence information can be mined by using a cyclic neural network model to generate a single representation for encoding the whole semantics of the sequence information, the invention adopts LSTM, and at a first-1 path, an LSTM layer inputs a spliced vector x l-1 which is an entity type, an entity value and a relation vector on the path, and the calculation mode is as follows:
A hidden state vector h l-1 is output, and the calculation mode is as follows:
zl＝tanh(Wzxl+Whhl-1+bz) (2)
fl＝σ(Wfxl+Whhl-1+bf) (3)
il＝σ(Wixl+Whhl-1+bi) (4)
ol＝σ(Woxl+Whhl-1+bo) (5)
cl＝fl⊙cl-1+il⊙zl (6)
hl＝ol⊙tanh(cl) (7)
Wherein c l∈Rd,z∈Rd represents a unit storage state vector and an information conversion module respectively, d' represents the number of hidden units, i l,ol and f l represent input, output and forget gates respectively, and W z,Wi,Wf,Wo∈Rd′×3d,Wh∈Rd′×d′ represents a mapping coefficient matrix; and b z,bi,bf and W o are bias vectors, σ (-) is an activation function, and as such, indicates the element-level product between the two vectors. The last output state vector h L can represent the entire path p K by using the memory mechanism.
Score prediction module:
The scoring prediction module is a set of paths P (u, i) = { P 1,p2,...,pK }, for a given user u and target item i, and u can reach i in G along a series of paths; output the prediction score of u interacting with i:
In the method, in the process of the invention, To evaluate the predicted outcome; f θ denotes a model with a parameter θ, which indicates whether user u interacted with item i is inferred from connectivity ρ (u, i).
And the generating recommendation module adopts two full-connection layers to project the final state into the output prediction scores. The predictive formula is:
W1, W2 represent the coefficient weights of layer 1 and layer 2, respectively, relu represents the activation function. Where τ= (u, act, i) represents a tuple; s (τ|p k)＝sk, representing the prediction score of the kth path from node u to node i, b representing the bias;
the generate recommendations module designs weighted pooling operations to aggregate the predicted scores of the interaction tuples:
Wherein the method comprises the steps of
Where γ is a hyper-parameter controlling each exponential weight, { s 1,s2,...,sK } represents the respective predictive score for the total K paths from node u to node i; such pooling can distinguish between different path importance.
Claims (4)
1. A multi-user recommendation system based on knowledge graph path reasoning is characterized in that: the system comprises a knowledge graph construction module, a path reasoning module and a scoring prediction module;
The knowledge graph construction module acquires user interaction historical data, constructs a knowledge graph G and transmits the knowledge graph G to the path reasoning module;
The path reasoning module generates a plurality of relation paths between the user and the target item according to the knowledge graph G and transmits the relation paths to the scoring prediction module;
The scoring prediction module performs scoring prediction on the received relation paths, and outputs a project recommendation list to a user according to the scoring prediction result;
The step of constructing the knowledge graph G includes:
a1 A) building a project association directed graph G 1; the entity of the item association directed graph G 1 comprises interaction items and candidate items, and the edge set comprises an association relationship between the entities;
a2 Constructing a directed graph G 2 of the interaction of the user with the item; the entities of the interactive directed graph G 2 include users and items; in the interaction directed graph G 2, if the interaction exists between the user and the item, an edge exists between the user and the item;
a3 Fusing the project association directed graph G 1 and the interaction directed graph G 2 of the user and the project to obtain a knowledge graph G;
project association directed graph G 1＝{(h,r,t)|h,t∈I1,r∈R1, obtained by modeling association data between candidate projects and interacted projects;
Wherein the tuple (h, r, t) indicates that there is a relationship r between the head node h and the tail node t; i 1 is a set of all candidate items; r 1 is a relationship set;
the directed graph G 2＝{(u,act,i)|u∈U,i∈I2,act∈R2 of the interaction between the user and the project is obtained by modeling the interaction record between the user and the project;
Wherein the tuple (u, act, i) represents that an interactive relationship act exists between the user u and the item i; And/> Respectively representing the user and the item set, M and N respectively representing the number of users and the number of items,/>R 2 represents an interaction relation set;
The knowledge graph g= { (h, R, t) |h, t E, R E R }, wherein set e=u U I 1, set r=r 1∪R2;
The step of generating a relationship path between the plurality of users and the target item by the path reasoning module comprises the following steps:
b1 Record the path between the user and the item on the knowledge graph G as the sequence of the entity and the relation Where e 1＝u,eL＝p;(el,rl,el+1) is the first tuple on path p, L-1 represents the number of all tuples on path p;
b2 For a given relationship path p k on the knowledge graph G, the path inference module segments each entity in the relationship path p k into a corresponding type and value, initializes the type and value into two vectors, respectively recorded as D represents the dimension size; the initial value of k is 1;
b3 Embedding a vector r l with the same dimension in a path p k, updating a relation path p 'k to [ e 1,r1,e2,...,el-1,rl-1 ], and enabling each element in the relation path p' k to represent an entity or a class of relation;
b4 Mining sequence information in the relationship path p' k using the recurrent neural network model stored on the path inference module, thereby generating an encoded path p k;
b5 Let k=k+1 and return to step b 2) until the encoded path set P (u, i) = { P 1,p2,...,pK };
The step of evaluating and predicting the received relation paths by the evaluation prediction module comprises the following steps:
c1 A) establishing a predictive function, namely:
In the method, in the process of the invention, To evaluate the predicted outcome; f θ denotes a model with a parameter θ, equation (1) indicates whether user u interacted with item i is determined by connectivity ρ (u, i);
c2 Calculating a state prediction value, namely:
Wherein W 1,W2 represents the coefficient weights of the 1 st full-connection layer and the 2 nd full-connection layer respectively; relu denotes an activation function; where τ= (u, act, i) represents a tuple; s (τ|p k)＝sk, representing the prediction score of the kth path from node u to node i, b representing the bias;
c3 Projecting the state prediction value into the prediction function to obtain an evaluation prediction result Namely:
Where σ (-) is the activation function, where the parameter g (s 1,s2,...,sK) is as follows:
Wherein γ is a hyper-parameter controlling each exponential weight; { s 1,s2,...,sK } represents the respective predictive scores of the K paths from node u to node i; s k represents the prediction score of the kth path from node u to node i.
2. The knowledge-graph-path-inference-based multi-user recommendation system of claim 1, wherein the recurrent neural network model is an LSTM neural network model, comprising an input layer, a hidden layer, and an output layer; the hidden layer comprises an input door, a forgetting door and an output door;
the input of the cyclic neural network model comprises entity types, entity values and relation vectors, and the output is an encoded path.
3. The multi-user recommendation system based on knowledge-graph path reasoning according to claim 2, wherein the input vector x l-1 and the output vector h L of the cyclic neural network model respectively satisfy the following formulas:
zl＝tanh(Wzxl+Whhl-1+bz) (6)
fl＝σ(Wfxl+Whhl-1+bf) (7)
il＝σ(Wixl+Whhl-1+bi) (8)
ol＝σ(Woxl+Whhl-1+bo) (9)
cl＝fl⊙cl-1+il⊙zl (10)
hl＝ol⊙tanh(cl) (11)
Wherein c l∈Rd,zl∈Rd represents a unit storage state vector and an information conversion module, and d' represents the number of hidden units; i l、ol and f l represent input, output and forget gates, respectively, and W z、Wi、Wf、Wo∈Rd'×3d,Wh∈Rd'×d' represents a mapping coefficient matrix; b z、bi、bf、bo is the bias vector; sigma (-) is an activation function, by which is meant the element-level product between the two vectors; e l-1、el'-1、rl-1 represents entity type, entity value, and relationship vector, respectively.
4. The multi-user recommendation system based on knowledge-graph path reasoning of claim 1, wherein: the items in the item recommendation list are predicted according to the scoresThe first N items in the ordering are decremented.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210583484.1A CN114820139B (en) | 2022-05-25 | 2022-05-25 | Multi-user recommendation system based on knowledge graph path reasoning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210583484.1A CN114820139B (en) | 2022-05-25 | 2022-05-25 | Multi-user recommendation system based on knowledge graph path reasoning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114820139A CN114820139A (en) | 2022-07-29 |
| CN114820139B true CN114820139B (en) | 2024-05-28 |
Family
ID=82519194
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210583484.1A Active CN114820139B (en) | 2022-05-25 | 2022-05-25 | Multi-user recommendation system based on knowledge graph path reasoning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114820139B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115618108B (en) * | 2022-10-24 | 2023-05-26 | 云南中烟工业有限责任公司 | Brand recommendation method based on knowledge graph in new retail model |
| CN116049438B (en) * | 2023-01-10 | 2023-06-02 | 昆明理工大学 | Knowledge graph-based group membership analysis method |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112417306A (en) * | 2020-12-10 | 2021-02-26 | 北京工业大学 | Method for optimizing performance of recommendation algorithm based on knowledge graph |
| CN112884548A (en) * | 2021-02-01 | 2021-06-01 | 北京三快在线科技有限公司 | Object recommendation method and device based on path reasoning and electronic equipment |
| CN113326384A (en) * | 2021-06-22 | 2021-08-31 | 四川大学 | Construction method of interpretable recommendation model based on knowledge graph |
| WO2021189971A1 (en) * | 2020-10-26 | 2021-09-30 | 平安科技（深圳）有限公司 | Medical plan recommendation system and method based on knowledge graph representation learning |
| CN114139066A (en) * | 2021-09-10 | 2022-03-04 | 重庆大学 | Collaborative filtering recommendation system based on graph neural network |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210406779A1 (en) * | 2020-06-26 | 2021-12-30 | Facebook, Inc. | System and methods for feature engineering based on graph learning |
| US20220058489A1 (en) * | 2020-08-19 | 2022-02-24 | The Toronto-Dominion Bank | Two-headed attention fused autoencoder for context-aware recommendation |
-
2022
- 2022-05-25 CN CN202210583484.1A patent/CN114820139B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2021189971A1 (en) * | 2020-10-26 | 2021-09-30 | 平安科技（深圳）有限公司 | Medical plan recommendation system and method based on knowledge graph representation learning |
| CN112417306A (en) * | 2020-12-10 | 2021-02-26 | 北京工业大学 | Method for optimizing performance of recommendation algorithm based on knowledge graph |
| CN112884548A (en) * | 2021-02-01 | 2021-06-01 | 北京三快在线科技有限公司 | Object recommendation method and device based on path reasoning and electronic equipment |
| CN113326384A (en) * | 2021-06-22 | 2021-08-31 | 四川大学 | Construction method of interpretable recommendation model based on knowledge graph |
| CN114139066A (en) * | 2021-09-10 | 2022-03-04 | 重庆大学 | Collaborative filtering recommendation system based on graph neural network |
Non-Patent Citations (1)
| Title |
|---|
| 基于知识图谱和Bi-LSTM的推荐算法;王钰蓥;《计算机与现代化》;20210915(第09期);90-98 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114820139A (en) | 2022-07-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114820139B (en) | Multi-user recommendation system based on knowledge graph path reasoning | |
| US11521221B2 (en) | Predictive modeling with entity representations computed from neural network models simultaneously trained on multiple tasks | |
| CN111881350B (en) | Recommendation method and system based on mixed graph structured modeling | |
| CN109190030B (en) | Implicit feedback recommendation method fusing node2vec and deep neural network | |
| CN111611472A (en) | Binding recommendation method and system based on graph convolution neural network | |
| CN103377262A (en) | Method and device for grouping users | |
| CN112989169B (en) | Target object identification method, information recommendation method, device, equipment and medium | |
| CN111523682B (en) | Method and device for training interactive prediction model and predicting interactive object | |
| CN115270001B (en) | Privacy protection recommendation method and system based on cloud collaborative learning | |
| Sha et al. | A Network‐Based Approach to Modeling and Predicting Product Coconsideration Relations | |
| CN112395515B (en) | Information recommendation method and device, computer equipment and storage medium | |
| CN115618108B (en) | Brand recommendation method based on knowledge graph in new retail model | |
| Abolghasemi et al. | A personality-aware group recommendation system based on pairwise preferences | |
| Alfantoukh et al. | Multi-stakeholder consensus decision-making framework based on trust: a generic framework | |
| Ahamed et al. | A recommender system based on deep neural network and matrix factorization for collaborative filtering | |
| Yu et al. | Collaborative group embedding and decision aggregation based on attentive influence of individual members: A group recommendation perspective | |
| CN115618128A (en) | Collaborative filtering recommendation system and method based on graph attention neural network | |
| Long | Three-dimensional-flow model of agent-based computational experiment for complex supply network evolution | |
| Gabbriellini et al. | A micro study on the evolution of arguments in amazon. com’s reviews | |
| Gao et al. | Recommendation system design for social media using reinforcement learning | |
| CN113468434B (en) | Resource recommendation method, device, readable medium and equipment | |
| CN112948668B (en) | Information recommendation method, electronic device and storage medium | |
| CN115391555A (en) | User-perceived knowledge map recommendation system and method | |
| CN115809339A (en) | Cross-domain recommendation method, system, device and storage medium | |
| CN115618124A (en) | Propagation popularity prediction method for bidirectional social influence learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |