CN112380434A - Interpretable recommendation system method fusing heterogeneous information network - Google Patents

Interpretable recommendation system method fusing heterogeneous information network Download PDF

Info

Publication number
CN112380434A
CN112380434A CN202011276253.3A CN202011276253A CN112380434A CN 112380434 A CN112380434 A CN 112380434A CN 202011276253 A CN202011276253 A CN 202011276253A CN 112380434 A CN112380434 A CN 112380434A
Authority
CN
China
Prior art keywords
user
attention
commodity
representation
information
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
Application number
CN202011276253.3A
Other languages
Chinese (zh)
Other versions
CN112380434B (en
Inventor
王英
贾天浩
王鑫
左万利
杨伟英
左祥麟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jilin University
Original Assignee
Jilin University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Jilin University filed Critical Jilin University
Priority to CN202011276253.3A priority Critical patent/CN112380434B/en
Publication of CN112380434A publication Critical patent/CN112380434A/en
Application granted granted Critical
Publication of CN112380434B publication Critical patent/CN112380434B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Landscapes

  • Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Development Economics (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to an interpretable recommendation system method fusing heterogeneous information networks, in particular to an offline recommendation system method based on a meta-path and deep neural network method. The method uses the similarity matrix generated by extracting the auxiliary information on the heterogeneous information network by the meta path as the supplementary information. And performing matrix decomposition on the similarity matrix by using a matrix decomposition method to obtain characteristic representations of a plurality of groups of users and commodities corresponding to the element paths. And distinguishing and combining the representations obtained based on different types of meta-paths by using an attention neural network, and finally combining the representation vectors by using a deep neural network to obtain a prediction score, and simultaneously selecting important meta-paths to generate an explanation.

Description

Interpretable recommendation system method fusing heterogeneous information network
Technical Field
The invention relates to an interpretable recommendation system method fusing heterogeneous information networks, in particular to an offline recommendation system method based on a meta-path and deep neural network method.
Background
Due to the rapid development of internet technology, people can obtain a large amount of online information such as news, movies, comments and the like on the internet, and the demand of users on the information in the information age is met. However, as the information grows explosively, users cannot obtain the information which is really useful for themselves when facing a large amount of information, the use efficiency of the information is reduced, and people face a serious information overload problem. The recommendation system is one of powerful methods for solving the information overload problem, and is a personalized information recommendation system for recommending information, products and the like interested by a user to the user according to the information requirements, interests and the like of the user. The recommendation system carries out personalized calculation by researching the interest preference of the user, and the system discovers the interest points of the user, thereby guiding the user to discover the own information requirement, providing personalized service for the user and establishing the close relationship with the user. Recommendation systems are now inseparable from people's daily lives and are also the primary tool of profit for user-oriented online services.
Recommendation systems have been widely used in many areas of online services, including search engines, e-commerce, online news and social media websites, among others.
The general paradigm of the recommendation method mainly includes three parts: data samples, representing learning and interaction methods. The data sample mainly refers to the interactive information of the user and the commodity and the additional characteristic information of the user commodity; the representation learning is to obtain potential low-dimensional representations of users and commodities according to input data; the interaction method measures the similarity of the potential representations of the user and the commodity to obtain a matching score. And then recommending the commodities with large matching scores to the user through matching score sorting. There are two main types of sample data for recommendation systems: explicit data and implicit data. Implicit data such as clicking and watching duration of a user, implicit data users do not express preferences of commodities clearly, but the implicit data is wide in existence and is paid more and more attention by recommendation system researchers. The display data such as the scores or praise of the goods by the user, etc. contains the determined preference information of the goods by the user, but the display data is relatively less and the acquisition cost is high. Explicit data general user scoring prediction tasks.
The key point of generating accurate, various and reasonable recommendation results is to obtain the characteristic vectors of users and commodities with rich representation capability, while the traditional recommendation system has the problem of sparsity of interactive information. One common idea for solving the data sparsity problem is to introduce some auxiliary information into the recommendation algorithm. The auxiliary information can enrich the description of the user and the article and enhance the mining capability of the recommendation algorithm, thereby effectively making up the sparseness or the deficiency of the interactive information. The heterogeneous information network contains rich attribute information and semantic association, and provides a potential auxiliary information source for the recommendation system. Potential high-order relations between users and commodities can be mined through different types of relation connection in a heterogeneous information network. The user information and the recommendation result can be connected based on the path, so that the satisfaction degree and the acceptance degree of the user on the recommendation result are improved, the trust of the user on a recommendation system is enhanced, and meanwhile, certain interpretability is generated.
Therefore, in a recommendation system fusing heterogeneous information, a recommendation system method capable of effectively fusing high-order information of a heterogeneous information network, and generating an accurate and effective recommendation result with the capability of effectively combining feature representations is needed.
Disclosure of Invention
The invention mainly aims to solve the problems of the conventional deep learning recommendation system method based on heterogeneous information;
the invention also aims to design an off-line recommendation system method which has higher accuracy, stronger interpretative capability and more effective recommendation effect;
it is yet another object of the present invention to solve two problems with the application of heterogeneous information networks to recommendation systems: aiming at the first problem, the invention aims to design a plurality of different types of meta-paths to be applied to a heterogeneous information network architecture to obtain a corresponding similarity matrix, and the meta-paths can be used as auxiliary information to make up the problem of sparse original user commodity interaction matrix, then potential feature representations of the user and the commodity are obtained through a representation learning method, aiming at the second problem, the invention designs an attention network for distinguishing the contribution degrees of the representations from different meta paths to the end user and commodity representations, and then aggregating the representations under the multi-element path based on the attention coefficient to obtain the final user and commodity representations, and finally obtaining the prediction scores of the user and commodity representations through a multi-layer perceptron.
The invention provides a recommendation system method for extracting heterogeneous network information based on meta-paths and combining feature representation by utilizing a deep neural network, which aims to solve the problems and achieve the aim.
The invention provides a recommendation system method based on meta-path and deep neural network, which comprises the following steps:
step one, obtaining a real data set containing rich heterogeneous information through the Internet;
step two, processing the data sets respectively, and extracting to obtain characteristic information of the commodity and the user and relation information between the characteristics;
thirdly, combining the obtained characteristic information and the relation information by using a meta-path method to obtain characteristic matrixes of the users and the commodities under different relations;
decomposing the obtained feature matrixes by a matrix decomposition method in the recommendation system to obtain user and commodity feature expression vectors under different element paths;
designing different attention neural networks to combine a plurality of feature expression vectors for the feature expression vectors of the users and the commodities under different element paths;
step six, obtaining the best attention neural network method through experimental result comparison, inputting the feature expression vectors obtained through the combination of the methods into a multilayer perceptron method to obtain the grade prediction of the user on the commodity;
and seventhly, based on the grade prediction of the user on the commodities, recommending commodities with higher grade prediction to the user, and giving reasonable recommendation reasons based on the attention coefficient in the attention network.
The specific process of acquiring and processing the heterogeneous information data in the first step and the second step is as follows:
firstly, acquiring a public recommendation system data set containing heterogeneous information through a website related to a recommendation system on the Internet;
secondly, extracting a plurality of similarity matrixes between every two entities according to different relations for entity characteristics, such as commodity comments, commodity descriptions, commodity types, commodity prices, commodity brands, purchasing relations and the like, contained in the heterogeneous information recommendation system data set, wherein the first two columns of the matrixes are names of the two entities, the third column is a score of the relation between the entities, the value of the score data is 1-5, the larger the score value is, the stronger the relation between the two entities is, and the smaller the score value is, the weaker the relation between the two entities is;
thirdly, the number of entities and the number of relations contained in the data set can be obtained through data processing, for example, Amazon data sets are taken as examples, the information contained in the Amazon data sets is counted as the following figure, the sparsity of the data sets also has differences, the sparsity of the data can be represented through density, and the data density is calculated according to the following formula:
Figure BDA0002779176610000041
where the numerator is the number of scores and the denominator is the product of the number of users and the number of items.
In the third and fourth steps, the specific process of combining the characteristic information and the relationship information by using the meta-path method and then performing matrix decomposition is as follows:
firstly, designing L meta-paths which are in a network mode TGA path defined as (a, R) having an originating node and a destination node, e.g.
Figure BDA0002779176610000042
AiE.g. A are different object types, Rie.R is the type of relationship between two object types, and the meta-path P can represent node A1And node AnThe complex relationship between the two or more of them,
Figure BDA00027791766100000511
one-path relation RiThe number of the data sets is the length of the meta path, and the meta path set for the data sets is shown in a figure below;
Figure BDA0002779176610000051
secondly, for the set meta-path, a corresponding user commodity similarity matrix can be obtained by adopting matrix multiplication, the initial node of the meta-path in the recommendation system is a user u, the target node is a commodity i, and the path in the Amazons data set is used as the path
Figure BDA0002779176610000052
For example, the path represents user u1Purchased commodity i1And goods i1And merchandise item i2Relationships belonging to the same class, similarity matrices based on meta-paths
Figure BDA0002779176610000053
Figure BDA0002779176610000054
Is a matrix multiplication operation between two relationship matrices,
Figure BDA00027791766100000512
representing entity type An-1With entity type AnA similarity matrix of the relationship between;
thirdly, after obtaining the user commodity similarity matrix corresponding to the L element paths, using matrix decomposition to obtain the potential representation of the user and the commodity, wherein the idea of the matrix decomposition is that one similarity matrix M can be decomposed into two low-rank matrices U and B, wherein U represents the potential characteristics of the user preference,b is a latent character contained in the commodity, and then passes
Figure BDA0002779176610000055
We can get a prediction similarity matrix
Figure BDA0002779176610000056
By optimizing the following objective function, the neural network parameters are continuously updated reversely, and finally the optimized objective function of potential vector representation and matrix decomposition of the user and the commodity can be obtained:
Figure BDA0002779176610000057
wherein
Figure BDA0002779176610000058
Is a prediction matrix obtained by multiplying a user and a commodity expression vector matrix,
Figure BDA0002779176610000059
and
Figure BDA00027791766100000510
is a dynamic parameter used to control the fourier norm regularization effect to avoid overfitting;
fourthly, for L (L is 8) element paths designed before, L similarity matrixes based on the element paths are obtained through entity relation matrix multiplication, and then the matrix decomposition operation is carried out simultaneously, so that L groups of different characteristic expressions U of users and commodities can be obtained(1)，B(1)，U(2)，B(2)…U(L)，B(L)。
In the fifth step, a specific process of designing different attention neural networks to combine a plurality of feature representations is as follows:
the present invention contemplates a method comprising two attention networks to combine the previously obtained L different sets of feature representations.
First, a local attention network is designed, wherein the aim of the local attention network is to obtain users combining L groups of different path information andcommodity representation ULAnd BLThe input of the local attention network is L groups of expression vectors obtained by matrix decomposition, and each group of expression vectors contains the user expression u corresponding to the pathiAnd commodity representation biFor L groups of users u is representediI e L, and inputting the i e L into the attention neural network facing to the user expression in sequence to obtain the corresponding uiAttention coefficient of (a)i：
DNNu(ui)＝Relu(Wu*(Relu(…(ui)…))+BU)
Figure BDA0002779176610000061
Wherein DNNu(ui) Is a user-oriented attention neural network, the input of which is user representation vectors of different paths to obtain corresponding matching numerical values, WUIs a neural network full-connection coefficient matrix, BUIs the bias of each layer, the activation function of each layer is Relu, and then the attention coefficient alpha is obtained by normalizing L output values of the neural network through the softmax functioniThe attention coefficient beta corresponding to the commodity representation under different paths can be obtained by adopting the same operation for the commodityi：
DNNb(bi)＝Relu(WB*(RElu(…(bi)…))+BB)
Figure BDA0002779176610000062
Then, according to the obtained attention coefficient alphaiAnd betaiUser and commodity representation U fusing L groups of different path information obtained by respectively combining user and commodity representation of L groups of pathsLAnd BLThe approach taken here is to multiply the user/product representation by the corresponding attention coefficient αi/βiThen directly splicing the L group alpha by a Concatenate operationi*ui/βi*bi：
UL＝Concate(α1*u1||α2*u2…||αi*ui)
BL＝Concate(β1*b1||β2*b2…||βi*bi)
User and commodity representation U obtained through local attention network layerLAnd BLThe information of different paths is contained, the information of the different paths is distinguished through an attention mechanism, the attention coefficient corresponding to the key path information which greatly contributes to the final recommendation prediction effect is large, the user representation information obtained through decomposition is reserved more, and finally, the user and commodity representation U is representedLAnd BLThe user and commodity interaction feature expression vector can be obtained by splicing, namely the input vector P of the multi-layer perceptron of the interaction prediction methodlocalAs follows:
Plocal＝Concate(UL||BL)
＝Concate(α1*u1||α2*u2…||αi*ui||β1*b1||β2*b2…||βi*bi)；
secondly, designing a global attention network, wherein the global attention network pays attention to distinguishing the contribution degree of the interactive feature representation corresponding to different paths, and firstly, representing vectors u corresponding to each group of pathsiAnd biSplicing to obtain the interactive feature representation p of the corresponding pathi，pi＝Concate(ui||bi) I ∈ L, then, set L into piSequentially inputting the attention neural network DNNz facing the interactive features to obtain corresponding attention coefficients thetai：
DNNz(pi)＝Relu(Wz*(Relu(…(pi)…))+Bz)
Figure BDA0002779176610000071
Where DNNz is a multi-layered neural network, the inputs are L sets of interactive feature representations piThe output is the corresponding match value. Wz、BzAnd Relu is a full connection coefficient matrix, a layer bias item and an activation function corresponding to the neural network respectively. Then, L output values of the neural network are normalized through a softmax function to obtain an attention coefficient thetaiThen, based on the obtained attention coefficient thetaiInteractive feature representation p combining L groups of pathsiObtaining user and commodity interactive feature representation P integrating different path informationglobalThe method adopted here is to represent the interactive characteristics of the L groups of paths as piBy the corresponding attention coefficient thetaiThen directly splicing the L group theta through a Concatenate operationi*pi：
Pglobal＝Concate(θ1*p1||θ2*p2…||θi*pi)
The interactive feature expression vectors of the user and the commodity obtained through the global attention network pay attention to the importance degree of the interactive feature information corresponding to different paths, not the importance degree of the user and the commodity corresponding to different paths paid attention to by local attention, and the interactive feature expression vectors are more suitable for recommendation tasks, wherein the interactive feature information can be used for mining deep information, and based on the attention coefficient in the global attention network, a recommendation result can be explained to a certain extent, namely, the path relation with a large corresponding attention coefficient is taken as a recommendation reason;
thirdly, designing an interactive method of entity vector representation, and obtaining an interactive feature representation P based on a local attention networklocalAnd a global attention network based interactive feature representation PglobalThen, the two methods need to be merged together so as to input the subsequent interaction method, and there are two methods in combination as follows:
P＝λ1*Plocal+λ2*Pglobal
P＝Concate(λ1*Plocal||λ2*Pglobal)
wherein λ1E (0, 1) and λ2E (0, 1) is a parameter for deciding the retention of each part of information, and the first method is to represent the local interactive feature by PlocalAnd global interactive feature representation PglobalThe second method is to splice and combine the two methods, based on the two methods, 2 method variants are designed in the experimental part, after the fused feature interactive representation is obtained, an interactive method is needed to combine the features for score prediction, most of the traditional methods adopt a decomposition machine method, the method has the advantages of simple operation and low calculation cost, but only can combine the feature relations of first order and second order, and the high-order feature relations are difficult to combine, so that the neural network is adopted as the feature interactive method, the neural network can automatically combine the high-order feature relations, the feature combination is not needed to be manually designed, a multilayer perceptron method is adopted here, the final feature interactive vector P is input, and the predicted score is output, as shown below:
ypred＝Relu(W*(Relu(…(P)…))+B)；
fourthly, designing a square loss function as an optimization target through a target task, namely a score prediction task of display data:
Loss＝(ypred-yreal)2+λ*||Para||2
wherein, ypredIs the score of the method prediction, yrealIs the user's score of the truth of the goods, ParaIs a trainable parameter in a neural network, the first part of the above equation is the squared difference of the true value and the predicted value, the second part is the L2 regularization, and the λ coefficient is used to control the regularization strength to prevent overfitting.
In the sixth step, the specific process of obtaining the optimal attention neural network method is as follows:
the first step, design 5 method variants, is the method (No attention) not using the attention mechanism, the method (Local attention) using only the Local attention network, the method (Global attention) using only the Global attention network, the method (Add attention) taking the addition operation when combining the Local interactive feature representation and the Global interactive feature representation and the method (coordinate attention) taking the splicing combination of the two;
secondly, performing comparison tests on the 5 variant methods, selecting an optimal method as a final prediction method according to results of the method variant comparison tests, and meanwhile, verifying whether the attention mechanism plays a role, and finding out how the attention mechanism improves the experimental accuracy and excavates a critical path;
and thirdly, determining an optimal attention neural network method, predicting by using the method to obtain the prediction score of the user on the commodity, and if the prediction score is closer to 5 points for a specific user and the commodity, indicating that the user likes the commodity more, and recommending the commodity with high prediction score to the user.
In the seventh step, a specific process for giving a reasonable recommendation reason by using the attention coefficient in the attention network is as follows:
firstly, obtaining and visualizing an attention coefficient of a global attention network, wherein each row in a visualization effect graph represents a prediction result, each column represents a meta path, one meta path in each record corresponds to one coefficient, a key path can be distinguished through the coefficient, and the larger the coefficient is, the larger the influence exerted by the meta path in the recommendation process is;
and secondly, selecting a meta path with the maximum coefficient as a reason for recommending the commodity to the user according to the obtained attention coefficient, wherein the meta path represents a piece of relation information, for example, the meta path UBCate shows that the reason for recommending the commodity to the user is because the user purchases the commodity of the same type as the commodity before.
Compared with the prior art, the invention has the beneficial effects that:
the off-line recommendation system method based on the meta-path and deep neural network method provided by the invention has the following beneficial effects;
the method comprises the steps of extracting a similarity matrix generated by auxiliary information on a heterogeneous information network by applying a meta-path to serve as a supplement of a scoring matrix, solving the problem of sparse original interaction data, carrying out matrix decomposition on the similarity matrix by applying a matrix decomposition method to obtain characteristic representations of a plurality of groups of users and commodities corresponding to the meta-path, distinguishing the influence of the representations obtained based on different types of meta-paths on final recommendation by using an attention mechanism, combining the representation vectors by using a deep neural network to obtain a prediction score, and simultaneously selecting an important meta-path to generate an explanation.
Drawings
FIG. 1 is a schematic diagram of the overall operation of the method of the present invention.
Fig. 2 is a schematic diagram of a process of extracting heterogeneous information by meta-path in the method of the present invention.
FIG. 3 is a schematic diagram of a process of matrix decomposition of a similarity matrix in the method of the present invention.
Fig. 4 is a process diagram of an attention neural network interaction method based on the method of the present invention.
Detailed Description
Please refer to fig. 1, fig. 2, fig. 3, and fig. 4:
example (b):
step one, acquiring two real data sets containing rich heterogeneous information through the Internet to process data sets, extracting and obtaining entity information of commodities and users and relationship information between the entities, wherein the first data set is Amazon, the data set comprises product comments and metadata from Amazon of a United states E-commerce website, the comment (rating, text, vote), main data (commodity description, commodity type, price, brand and image characteristics) and purchasing relationship, the second data set is Yelp, a general data set, an internal data set disclosed by Yelp of the United states commenting website, the Yelp data set is a subset of covered merchant, rating and user data, the subset comprises user rating, merchant information, city information, merchant attributes and user characteristics, and the Amazon data set comprises 195791 pieces of rating data of 6170 users and 2753 pieces of commodities, the Yelp dataset contained 198397 scoring data for 16239 users and 14284 items, with values for scoring data ranging from 1-5:
Figure BDA0002779176610000111
extracting relationship information into a matrix, wherein each row [ Business, Category, value ], the first two columns are entities, and the last column is a relationship between the two entities, such as [2200,5,1], which indicates that the 2200 number commodity belongs to the 5 th class;
step three, for the obtained characteristic information and relation information, combining by using a meta-path method to obtain a characteristic matrix of the user and the commodity under different relations, for example, the meta-path U → B ← U → B used by the data set Amazon indicates that the commodity purchased by the user with the same preference is recommended to the user, which can be regarded as a collaborative filtering model based on the user, and a similarity matrix corresponding to the meta-path can be obtained by a formula
Figure BDA0002779176610000121
The meta-path adopted by the invention is obtained by the following diagram:
Figure BDA0002779176610000122
and step four, decomposing the plurality of feature matrixes obtained by the method of matrix decomposition in the recommendation system to obtain the feature expression vectors of the users and the commodities under different element paths, wherein the feature vector U with the user ID of 32010 and the feature vector v with the commodity ID of 2420 obtained under the path U → B ← U → B are respectively as follows: u ═ (-1.032,0.256,0.321,1.056,0.985,0.963,0.942,0.456,0.587, -0.109, -0.746, -0.124,0.872,0.643,0.519, -0.467), v ═ (0.532,0.756, -0.721,0.161,0.211,0.473, -0.538, -0.743,1.012,1.051, -0.851, -0.732,0.463,0.742,0.658, -0.671);
step five, for the user and commodity feature representation vectors under different element paths, designing different attention neural networks to combine a plurality of feature representation vectors, taking the user 32010 and the commodity 2420 as examples, for the local attention network, the attention coefficients under different element paths are (0.42,0.10, 0.08, 0.22,0.15,0.03), respectively, representing the influence of different element paths on the representation vectors, so that the vector representation combining the plurality of element paths can be a U-0.42U 8 + 0.10U 2+ 0.08U 3+ 0.22U 8 + 0.15U 5+ 0.03U 6, V-0.42V 1+ 0.10V 2+ 0.84 + 0.22U 4642V + 8442V + 8428, V + 7V 94V 42 + 0.7V + 0.42V + V94V + 3V + 0.42V + 3V + V42V +0 g 48 + V +0 g 12V +0 g 12V + 3V + 3V + 12V + 3V +, first, each element path u is spliced, v is obtained as p, (u, v), then p is obtained as attention coefficients (0.4,0.11, 0.09, 0.20,0.16,0.04) under different element paths, and Pg is obtained as 0.4 u1+0.11 u2+0.09 u3+0.2 u4+0.16 u5+0.04 u6, 0.4 v1+0.11 v2+0.09 v3+0.2 v4+0.16 v5+0.04 v6, where the combination of Pl and Pg is input into the interaction model in a second mode (L1, L2);
step six, obtaining the best attention neural network method through experimental result comparison, inputting the feature expression vectors obtained through the method combination into a multilayer perceptron method to obtain the grade prediction of the user on the commodity, for example, if the prediction grades of the user 32010 on the commodities 124, 563, 2498, 4167 and 8511 are respectively 1, 2, 3, 4 and 5, the user is shown to probably like the commodity 8511, so that the commodity 8511 can be recommended to the user;
step seven, based on the rating prediction of the user on the commodity, the commodity with higher rating prediction can be recommended to the user, and a reasonable recommendation reason is given based on the attention coefficient in the attention network, taking the rating record No. 5 in the visual graph as an example, the user ID of the commodity is 2840, the commodity ID is 8480, the rating is 5, the maximum corresponding attention coefficient is C, namely the path ubcate b, and the explanation can be provided based on the path: the recommendation of the item 8480 to the user 2840 is due to the user having previously purchased the same kind of item as the item.

Claims (1)

1. An interpretable recommendation system method fusing heterogeneous information networks is as follows:
(1) obtaining a real data set containing rich heterogeneous information through the Internet, respectively processing the data sets, extracting to obtain characteristic information of the commodity and the user and relation information between the characteristics:
1) acquiring a public recommendation system data set containing heterogeneous information through a website related to a recommendation system on the Internet;
2) for entity characteristics contained in the heterogeneous information recommendation system data set, such as comments of commodities, description of commodities, commodity types, commodity prices, commodity brands, purchasing relations and the like, a plurality of similarity matrixes between every two entities are extracted according to different relations, the first two columns of the matrixes are names of the two entities, and the third column is a score of the relation between the entities;
3) the entity number and the relation number contained in the data set can be obtained through data processing; the sparsity of the data set can be represented by the density, which is calculated as follows:
Figure FDA0002779176600000011
(2) for the obtained characteristic information and the relation information, combining by using a meta-path method to obtain similarity matrixes of the users and the commodities under different relations, and decomposing the obtained multiple similarity matrixes by using a matrix decomposition method in a recommendation system to obtain characteristic expression vectors of the users and the commodities under different meta-paths:
1) firstly, designing L element paths;
2) for the set meta path, the corresponding user commodity similarity matrix can be obtained by adopting matrix multiplication, and the similarity matrix based on the meta path
Figure FDA0002779176600000012
Figure FDA0002779176600000013
Is a matrix multiplication operation between two relationship matrices,
Figure FDA0002779176600000014
representing entity type An-1And entity typeAnA similarity matrix of the relationship between;
3) after a user commodity similarity matrix corresponding to the L element paths is obtained, potential representations of the user and the commodity are obtained by matrix decomposition, and an optimization objective function of the matrix decomposition is as follows:
Figure FDA0002779176600000021
wherein
Figure FDA0002779176600000022
Is a prediction matrix obtained by multiplying a user and a commodity expression vector matrix,
Figure FDA0002779176600000023
and
Figure FDA0002779176600000024
is a dynamic parameter used to control the fourier norm regularization effect to avoid overfitting;
4) for the previously designed L element paths, L similarity matrixes based on the element paths are obtained through entity relationship matrix multiplication, and the matrix decomposition operation is carried out simultaneously to obtain L groups of characteristic representation U of different users and commodities(1)，B(1)，U(2)，B(2)…U(L)，B(L)；
(3) For the feature expression vectors of users and commodities under different element paths, different attention neural networks are designed to combine a plurality of feature expression vectors:
1) designing a local attention network, wherein the aim of the local attention network is to obtain a user and commodity representation U combining L groups of different path informationLAnd BLFor L groups of users u is representediI e L, and inputting the i e L into the attention neural network facing to the user expression in sequence to obtain the corresponding uiAttention coefficient of (a)i：
DNNu(ui)＝Relu(WU*(Relu(…(ui)…))+BU)
Figure FDA0002779176600000025
Wherein DNNu(ui) Is a user-oriented attention neural network whose input is a user representation vector of different paths, WUIs a neural network full-connection coefficient matrix, BUIs the bias of each layer, the activation function of each layer is Relu; then, L output values of the neural network are normalized through a softmax function to obtain an attention coefficient alphaiThe attention coefficient beta corresponding to the commodity representation under different paths can be obtained by adopting the same operation for the commodityi：
DNNb(bi)＝Relu(WB*(Relu(…(bi)…))+BB)
Figure FDA0002779176600000026
Then, according to the obtained attention coefficient alphaiAnd betaiUser and commodity representation U for combining different path information of L groups by combining user and commodity representation of L groups of paths respectivelyLAnd BLMultiplying the user/product representation by the corresponding attention coefficient alphai/βiThen Concatenate operation directly splices L group alphai*ui/βi*bi：
UL＝Concate(α1*u1||α2*u2…||αi*ui)
BL＝Concate(β1*b1||β2*b2…||βi*bi)
Finally, by representing the user and the goods as ULAnd BLThe user and commodity interaction feature expression vector can be obtained by splicing, namely the input vector P of the multi-layer perceptron of the interaction prediction methodlocalAs follows:
Plocal＝Concate(UL||BL)
＝Concate(α1*u1||α2*u2…||αi*ui||β1*b1||β2*b2…||βi*bi)；
2) designing a global attention network, firstly, representing vectors u corresponding to each group of pathsiAnd biSplicing to obtain the interactive feature representation p of the corresponding pathi，pi＝Concate(ui||bi) I ∈ L, then set L into piSequentially inputting attention neural network DNN oriented to interactive featuresZIn the method, a corresponding attention coefficient theta is obtainedi：
DNNZ(pi)＝Relu(WZ*(Relu(…(pi)…))+BZ)
Figure FDA0002779176600000031
Wherein DNNZIs a multi-layer neural network, and the input is an interactive feature representation p of the L groupsi，WZ、BZRelu is a full-connection coefficient matrix, a layer bias item and an activation function corresponding to the neural network respectively, and then L output values of the neural network are normalized through a softmax function to obtain an attention coefficient thetai；
According to the obtained attention coefficient thetaiInteractive feature representation p combining L groups of pathsiObtaining user and commodity interactive feature representation P combining different path informationglobalExpressing the interactive characteristics of the L groups of paths as piBy the corresponding attention coefficient thetaiThen directly splicing the L group theta through a Concatenate operationi*pi：
Pglobal＝Concate(θ1*p1||θ2*p2…||θi*pi)；
3) Designing an interactive method of entity vector representation, and representing interactive features based on local attention networkPlocalAnd a global attention network based interactive feature representation PglobalThe combination method adopted by combining the components comprises the following two methods:
P＝λ1*Plocal+λ2*Pglobal
P＝Concate(λ1*Plocal||λ2*Pglobal)
wherein λ1E (0, 1) and λ2E (0, 1) is a parameter for deciding the retention of each part of information, and the first method is to represent the local interactive feature by PlocalAnd global interactive feature representation PglobalAdding, and the second method is to splice and combine the two;
the neural network is adopted as a feature interaction method, so that the high-order feature relationship can be automatically combined, a final feature interaction vector P is input, and a predicted score is output, wherein the predicted score is as follows:
ypred＝Relu(W*(Relu(…(P)…))+B)；
4) designing a square loss function as an optimization target through a target task, namely a score prediction task of display data:
Loss＝(ypred-yreaal)2+λ*||Para||2
wherein, ypredIs the score of the method prediction, yrealThe score of a user on the truth of the commodity, Para is a trainable parameter in a neural network, the first part of the above formula is the square difference between a truth value and a predicted value, the second part is L2 regularization, and a lambda coefficient is used for controlling the regularization strength to prevent overfitting;
(4) and obtaining the best attention neural network method through experimental result comparison, inputting the feature expression vectors obtained by the method combination into a multilayer perceptron method to obtain the grade prediction of the user on the commodity:
1) designing 5 variant methods, namely a method without using an attention mechanism, a method only using a local attention network, a method only using a global attention network, a method adopting addition operation when combining a local interactive feature representation and a global interactive feature representation and a method adopting splicing and combining the local interactive feature representation and the global interactive feature representation;
2) comparative tests were performed on the above 5 variant methods; selecting an optimal method as a final prediction method through the result of the method variant comparison test;
(5) based on the score prediction of the user on the commodities, the commodity with higher score prediction can be recommended to the user, and reasonable recommendation reasons are given based on the attention coefficient in the attention network:
1) determining an optimal attention neural network method, predicting by using the method to obtain the prediction score of a user on commodities, and recommending the goods with high prediction score to the user;
2) obtaining and visualizing an attention coefficient of the global attention network;
3) based on the obtained attention coefficient, an explanation is generated based on the meta path having the largest coefficient as a reason for recommending the commodity to the user.
CN202011276253.3A 2020-11-16 2020-11-16 Interpretable recommendation method fusing heterogeneous information network Active CN112380434B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011276253.3A CN112380434B (en) 2020-11-16 2020-11-16 Interpretable recommendation method fusing heterogeneous information network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011276253.3A CN112380434B (en) 2020-11-16 2020-11-16 Interpretable recommendation method fusing heterogeneous information network

Publications (2)

Publication Number Publication Date
CN112380434A true CN112380434A (en) 2021-02-19
CN112380434B CN112380434B (en) 2022-09-16

Family

ID=74584318

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011276253.3A Active CN112380434B (en) 2020-11-16 2020-11-16 Interpretable recommendation method fusing heterogeneous information network

Country Status (1)

Country Link
CN (1) CN112380434B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114036837A (en) * 2021-11-08 2022-02-11 中国人民解放军国防科技大学 Equipment combination method, system, equipment and storage medium based on co-construction sharing
CN114048396A (en) * 2022-01-13 2022-02-15 湖南大学 NQI service recommendation method and system based on heterogeneous information network and deep learning
CN114519097A (en) * 2022-04-21 2022-05-20 宁波大学 Academic paper recommendation method for heterogeneous information network enhancement
CN114637888A (en) * 2022-05-18 2022-06-17 深圳市华曦达科技股份有限公司 Video pushing method and device

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105761154A (en) * 2016-04-11 2016-07-13 北京邮电大学 Socialized recommendation method and device
CN105913125A (en) * 2016-04-12 2016-08-31 北京邮电大学 Heterogeneous information network element determining method, link prediction method, heterogeneous information network element determining device and link prediction device
CN106503028A (en) * 2015-09-08 2017-03-15 北京邮电大学 Recommend method and system
US20180341710A1 (en) * 2017-05-24 2018-11-29 Microsoft Technology Licensing, Llc Heterogeneous social networking
CN110751284A (en) * 2019-06-06 2020-02-04 北京嘀嘀无限科技发展有限公司 Heterogeneous information network embedding method and device, electronic equipment and storage medium

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106503028A (en) * 2015-09-08 2017-03-15 北京邮电大学 Recommend method and system
CN105761154A (en) * 2016-04-11 2016-07-13 北京邮电大学 Socialized recommendation method and device
CN105913125A (en) * 2016-04-12 2016-08-31 北京邮电大学 Heterogeneous information network element determining method, link prediction method, heterogeneous information network element determining device and link prediction device
US20180341710A1 (en) * 2017-05-24 2018-11-29 Microsoft Technology Licensing, Llc Heterogeneous social networking
CN110751284A (en) * 2019-06-06 2020-02-04 北京嘀嘀无限科技发展有限公司 Heterogeneous information network embedding method and device, electronic equipment and storage medium

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
M.JAMALI等: "HeteroMF: recommendation in heterogeneous information networks using context dependent factor models", 《WWW "13: PROCEEDINGS OF THE 22ND INTERNATIONAL CONFERENCE ON WORLD WIDE WEB》 *
XIN WANG等: "Attention-Guide Walk Model in Heterogeneous Information Network for Multi-Style Recommendation Explanation", 《PROCEEDINGS OF THE AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE》 *
刘露等: "异质网络中基于节点影响力的相似度度量方法", 《电子学报》 *
吴浩等: "基于异质用户网络嵌入的服务推荐方法研究", 《计算机工程与科学》 *
李亚莹: "基于异构信息网络表征学习的推荐方法", 《现代计算机》 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114036837A (en) * 2021-11-08 2022-02-11 中国人民解放军国防科技大学 Equipment combination method, system, equipment and storage medium based on co-construction sharing
CN114036837B (en) * 2021-11-08 2024-06-04 中国人民解放军国防科技大学 Equipment combination method, system, equipment and storage medium based on co-construction sharing
CN114048396A (en) * 2022-01-13 2022-02-15 湖南大学 NQI service recommendation method and system based on heterogeneous information network and deep learning
CN114048396B (en) * 2022-01-13 2022-03-18 湖南大学 NQI service recommendation method and system based on heterogeneous information network and deep learning
CN114519097A (en) * 2022-04-21 2022-05-20 宁波大学 Academic paper recommendation method for heterogeneous information network enhancement
CN114637888A (en) * 2022-05-18 2022-06-17 深圳市华曦达科技股份有限公司 Video pushing method and device
CN114637888B (en) * 2022-05-18 2022-08-02 深圳市华曦达科技股份有限公司 Video pushing method and device

Also Published As

Publication number Publication date
CN112380434B (en) 2022-09-16

Similar Documents

Publication Publication Date Title
CN112380434B (en) Interpretable recommendation method fusing heterogeneous information network
Zhang et al. The study of the effect of online review on purchase behavior Comparing the two research methods
Singh et al. Quantifying the effect of eWOM embedded consumer perceptions on sales: An integrated aspect-level sentiment analysis and panel data modeling approach
Bi et al. A deep neural networks based recommendation algorithm using user and item basic data
Hildebrandt et al. A recommender system for complex real-world applications with nonlinear dependencies and knowledge graph context
Van Dat et al. Solving distribution problems in content-based recommendation system with gaussian mixture model
Wang et al. Joint representation learning with ratings and reviews for recommendation
Ge et al. Helpfulness-aware review based neural recommendation
Alamdari et al. An image-based product recommendation for E-commerce applications using convolutional neural networks
Won et al. Perceptual mapping based on web search queries and consumer forum comments
Qian et al. Why some products compete and others don't: A competitive attribution model from customer perspective
Yengikand et al. DHSIRS: a novel deep hybrid side information-based recommender system
Ibrahim et al. Improved Hybrid Deep Collaborative Filtering Approach for True Recommendations.
Li et al. The role of sentiment tendency in affecting review helpfulness for durable products: nonlinearity and complementarity
Wang et al. Data science for next-generation recommender systems
Li et al. Impact of data-driven online financial consumption on supply chain services
Yang et al. A novel customer-oriented recommendation system for paid knowledge products
Almutairi et al. Review helpfulness prediction: Survey
Lu Knowledge distillation-enhanced multitask framework for recommendation
Nagy et al. A review for recommender system models and deep learning
Padti et al. Hybrid Movie Recommender System
Yin et al. An interpretable neural network TV program recommendation based on SHAP
Wang et al. NAUI: Neural attentive user interest model for cross-domain CTR prediction
Chandrasekaran et al. An intelligent framework for online product recommendation using collaborative filtering
Iswari et al. User-Generated Content Extraction: A Bibliometric Analysis of the Research Literature (2007–2022)

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