CN114238758A - User portrait prediction method based on multi-source cross-border data fusion - Google Patents

User portrait prediction method based on multi-source cross-border data fusion Download PDF

Info

Publication number
CN114238758A
CN114238758A CN202111531109.4A CN202111531109A CN114238758A CN 114238758 A CN114238758 A CN 114238758A CN 202111531109 A CN202111531109 A CN 202111531109A CN 114238758 A CN114238758 A CN 114238758A
Authority
CN
China
Prior art keywords
user
vector
representing
output
interaction
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
CN202111531109.4A
Other languages
Chinese (zh)
Other versions
CN114238758B (en
Inventor
周仁杰
郭星宇
张纪林
万健
刘畅
赵乃良
殷昱煜
蒋从锋
刘焱
李炳
陈青雯
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.)
Zhejiang Panshi Information Technology Co ltd
Hangzhou Dianzi University
Original Assignee
Zhejiang Panshi Information Technology Co ltd
Hangzhou Dianzi 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 Zhejiang Panshi Information Technology Co ltd, Hangzhou Dianzi University filed Critical Zhejiang Panshi Information Technology Co ltd
Priority to CN202111531109.4A priority Critical patent/CN114238758B/en
Publication of CN114238758A publication Critical patent/CN114238758A/en
Application granted granted Critical
Publication of CN114238758B publication Critical patent/CN114238758B/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
    • 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
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Animal Behavior & Ethology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a user portrait prediction method based on multi-source cross-border data fusion, and aims to solve the problem of inaccurate user characteristic prediction caused by sparse project characteristics, high-order structural characteristic loss and user behavior sequence characteristic loss in the prior art. Based on E-commerce data generated by a user, commodity characteristics are expanded by using a knowledge graph, historical purchase records of the user are fully mined by using a graph convolution network, potential purchase characteristics of the user are predicted by using a recurrent neural network, and accuracy of user portrait prediction is effectively improved. The method has the advantages that the problem of sparse commodity features is solved through the knowledge graph, the problem of high-order structural feature loss is solved through the graph convolution neural network, the problem of user behavior sequence feature loss is solved through the recurrent neural network, and a good foundation is laid for improving the performance of a recommendation system.

Description

User portrait prediction method based on multi-source cross-border data fusion
Technical Field
The invention relates to a user portrait prediction method based on multi-source cross-border data fusion, which is constructed according to a historical order sequence and content information of user shopping.
Background
Along with the development of internet technology and intelligent devices, various mobile applications are emerging and penetrating the lives of people, and the generated information is also increasing explosively. This makes it difficult for people to efficiently obtain the desired information, while enterprises have difficulty in accurately pushing products, information, etc. to users. The recommendation system is based on the user portrait, the user portrait is efficiently constructed, and the enterprise can realize fine marketing and accurate recommendation.
Online shopping has become an extremely common thing in today's life, and many users can directly or indirectly provide personal information to a shopping platform while enjoying the convenience of online shopping. The direct information includes sex, age, place of residence, etc., and the indirect information includes browsing record, purchasing record, collecting record, etc. According to the information of the user, the shopping platform can construct a virtual portrait of the user in the internet, so that the needed commodities can be accurately recommended to the user, and the benefit of the shopping platform is improved.
The technology for viewing user portraits is mainly applied to the field of personalized recommendation, and various methods for predicting user portraits labels comprise SVM, decision trees, LR and other traditional shallow learning models with good effect. However, as data generated by users in a big data background is explosively increased and feature dimensions are increased, the limitation of the flattened structure of the traditional shallow learning model begins to be highlighted. For example, in typical problems of user click rate estimation, conversion rate estimation and the like, the processed features as input have the characteristics of high latitude and high sparsity, and the traditional shallow learning method faces certain challenges in user label prediction because complex nonlinear relations among the features cannot be found.
In the field of electronic commerce, the historical purchasing behavior of the user contains the behavior information of the user. The accuracy of user portrayal can be effectively improved through the purchase records of the user, and the performance of a recommendation system is further improved. For example, if a user's purchase history includes a large amount of "Hua's" brand, indicating that the user is "pollen", the user will not buy the phone with a high probability if the recommendation system recommends the "iphone" brand of phone to the user. And if the recommendation system pushes the newly released 'Huayi' mobile phone to the user, the user may buy the mobile phone if the user needs to change the phone. The so-called "Huacheng" and "iphone" are implicit features hidden in the historical purchasing behavior of the user. Other implicit features such as "efficacy", "genre", "price", "speaker", etc. of the product or "director", "producer", "genre", etc. of the movie. The implicit characteristics of the items often have sparsity problems in the network platform. In addition, most of the above methods do not mine the association between users and between projects, and most of the methods use user feature prediction as a classification task, and each feature of the user is relatively independent, so that the associated features between users and between projects are lost to a certain extent, and the representation vector of one user cannot be effectively learned to be used as the user feature prediction.
The invention utilizes the knowledge graph to supplement the characteristics of the user historical purchased commodities and provides a user portrait prediction method for learning the high-order structural characteristics of the user based on the graph convolution neural network. Meanwhile, the characteristics of the user are supplemented by the recurrent neural network according to the historical sequence of the purchase order of the user, and a complete user portrait prediction method based on multi-source cross-border data fusion is constructed.
Disclosure of Invention
The invention aims to solve the problems of project feature sparseness, high-order structural feature loss and user behavior sequence feature loss in the prior art, and provides a user portrait prediction method based on multi-source cross-border data fusion.
The technical scheme adopted by the invention is as follows:
step 1: collecting information generated by interaction of a user on a shopping platform;
step 2: constructing a heterogeneous knowledge graph and a user historical interaction sequence;
and step 3: constructing an embedded matrix;
and 4, step 4: constructing a user portrait prediction model of multi-source cross-border data fusion, training, and obtaining an optimal parameter model after model parameters are converged;
and 5: and (4) predicting user characteristics by using the user portrait prediction model for constructing multi-source cross-boundary data fusion obtained in the step (4).
The invention also aims to provide a user portrait prediction device based on multi-source cross-boundary data fusion, which comprises a memory, a processor and a sequence perception and image convolution based neural network model program stored on the memory and capable of running on the processor, wherein the sequence perception and image convolution based neural network model program realizes the steps of the user portrait prediction method based on multi-source cross-boundary data fusion when being executed by the processor.
It is still another object of the present invention to provide a storage medium storing a multi-source cross-boundary data fused user portrait prediction model program, which when executed by a processor implements the steps of the above-mentioned multi-source cross-boundary data fused user portrait prediction method.
The technical scheme provided by the invention has the following beneficial effects:
(1) according to the historical orders of the users, the commodity characteristics are expanded by adopting the knowledge graph, so that the problem that the commodity characteristics in E-commerce data are scarce is solved;
(2) constructing a knowledge subgraph by using the commodities and the related knowledge map triples; the method comprises the steps of fully learning knowledge subgraph node characteristics by using a graph convolution network, keeping the structural characteristics of a graph as much as possible, avoiding characteristic loss caused by a training process, and obtaining a representation vector capable of fully representing an entity and local neighbor characteristics of the entity; the problem of high-order structural feature loss is solved;
(3) and extracting features hidden in the user behavior sequence by using a recurrent neural network for the historical order sequence of the user. The high-order structural characteristics obtained by combining the learning of the graph convolution network model solve the problem of user behavior sequence characteristic loss, and further improve the user portrait prediction capability of the model.
Drawings
FIG. 1 is a flow chart according to the present invention;
FIG. 2 is a diagram of a model structure;
FIG. 3 is a schematic diagram of a heterogeneous knowledge graph;
Detailed Description
Embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
A specific flow description of a user portrait prediction method of multi-source cross-border data fusion is shown in FIG. 1, in which:
step 1: and collecting information generated by interaction of the user on the shopping platform.
The information collected includes:
(1) basic information of the user includes gender and age.
(2) And the user behavior records comprise the time of purchasing the commodity, the commodity number, the commodity name and the like.
Step 2: constructing a heterogeneous knowledge graph and a user historical interaction sequence;
2-1 construction of heterogeneous knowledge graph
2-1-1, performing word segmentation on the commodity name to obtain a word segmentation result set { i }1，i2，...，im，...}，imRepresenting the mth participle;
2-1-2, performing 2-round recursion search on the segmentation result set in the public knowledge graph, discarding segmentation results which do not exist in the public knowledge graph, and forming an entity set epsilon (e) by using the remaining segmentation in the segmentation result set and entities searched in the public knowledge graph1，e2，...，en,., and further constructing them into triplets (i)m，contain，en) Cotain represents imAnd enThe incidence relation between the two; using the above triplet (i)m，contain，en) Constructing knowledge subgraphs corresponding to commodity names
Figure BDA0003410781700000041
2-1-3 knowledge subgraph corresponding to commodity name
Figure BDA0003410781700000042
Integrated into a heterogeneous knowledge graph
Figure BDA0003410781700000043
The heterogeneous knowledge graph
Figure BDA0003410781700000044
Includes node V and edge E; the node V comprises a user set U, a commodity name set I and an entity set epsilon; the edges include three types, respectively entity-entity intellectual graph relationships
Figure BDA0003410781700000045
Commodity name-user interaction record EiuAnd the users have the same click behavior between pairs Euu；
The entity-entity knowledge graph relationship
Figure BDA0003410781700000046
Is the relationship between any two entities in the entity set;
2-2, constructing a user-commodity name interaction matrix according to the user set, the commodity name set and the commodity name-user interaction records
Figure BDA0003410781700000047
N represents the number of users, and M represents the number of names of commodities;
y in the user-merchandise name interaction matrixuv1 indicates that the user u interacts with the product name v (e.g., purchases, browses, clicks, etc.), if yuv0 means that user u has not made an interaction with the trade name v;
2-3, further constructing a user history interaction sequence set according to the user-commodity name interaction matrix, wherein the set comprises the following steps:
Figure BDA0003410781700000051
wherein
Figure BDA0003410781700000052
The name of the item representing the user u's historical interaction at the ith time,
Figure BDA0003410781700000053
represents user u and
Figure BDA0003410781700000054
the moment at which the interaction occurs;
and step 3: according to the knowledge graph relation among the user set, the entity set epsilon and the entity-entity
Figure BDA0003410781700000055
Further constructing a user embedding matrix
Figure BDA0003410781700000056
Entity embedded matrix
Figure BDA0003410781700000057
And a user adjacency matrix
Figure BDA0003410781700000058
Wherein D represents a dimension of the vector; each element in the user adjacency matrix represents the similarity of the click behaviors of two users;
user adjacency matrix element
Figure BDA0003410781700000059
Representing user u1With user u2With a similar click-through behavior, the user may,
Figure BDA00034107817000000510
Figure BDA00034107817000000511
representing user u1With user u2There is no similar click behavior;
and 4, step 4: constructing a user portrait prediction model of multi-source cross-border data fusion;
the user portrait prediction model of multi-source cross-border data fusion comprises an input embedding layer, a heterogeneous knowledge graph convolution layer, a user behavior sequence perception layer and an output layer:
4-1 input embedding layer:constructing a set of user interaction entities N using a set of historical interaction sequences of userse(u); vectorizing user with embedded matrix, vectorizing user interaction entity with embedded matrix, and displayingi(u) acquiring the user' S neighbor embedded vector according to the user adjacency matrix, and constructing a neighbor user set Su(u)；
4-2 heterogeneous knowledge map convolutional layer: after the representation vector of the user interaction entity enters the heterogeneous knowledge graph convolutional layer, two parts of operations are executed;
the 4-2-1 user interaction entity obtains a user-commodity name expression vector with neighbor features through H-round iterative aggregation of neighbor topological structure features
Figure BDA00034107817000000512
4-2-2 user u's neighboring user set expression vector and user u expression vector are aggregated to obtain user neighboring feature expression vector
Figure BDA00034107817000000513
4-2-3 user-Commodity name representation vector
Figure BDA00034107817000000514
And user neighbor feature representation vector
Figure BDA00034107817000000515
Adding the spliced vector and the user u representation vector to obtain an output vector of the heterogeneous knowledge graph convolutional layer
Figure BDA00034107817000000516
4-3, the user behavior sequence perception layer adopts LSTM or GRU to model user sequence characteristics so as to extract potential interest of users; to be provided with
Figure BDA0003410781700000061
Obtaining a vector with the same dimensionality as the output of the heterogeneous knowledge map convolutional layer for input;
the first method is as follows: user sequence feature modeling using LSTM
Hiding state of last moment of recurrent neural network
Figure BDA0003410781700000062
And cell status
Figure BDA0003410781700000063
Adding to obtain the output vector of the LSTM module:
Figure BDA0003410781700000064
wherein the content of the first and second substances,
Figure BDA0003410781700000065
representing historical interaction sequences of user u
Figure BDA0003410781700000066
The output vector after being processed by the LSTM module,
Figure BDA0003410781700000067
representing the cell state output by the LSTM neural network at the last moment,
Figure BDA0003410781700000068
representing the hidden state of the LSTM neural network output at the last time, T representing the last time,
Figure BDA0003410781700000069
represents an addition at the element level;
and then, carrying out spatial transformation on the output vector of the LSTM module, and converting the output vector into a user behavior sequence representation vector with the same dimension as the user representation vector:
Figure BDA00034107817000000610
wherein the content of the first and second substances,
Figure BDA00034107817000000611
the sequence of behaviors representing user u represents a vector,
Figure BDA00034107817000000612
and
Figure BDA00034107817000000613
respectively representing a weight matrix and a bias for spatial transformation, wherein P represents the number of LSTM hidden layer neurons;
the second method comprises the following steps: user sequence feature modeling using GRUs
Hidden state of last minute
Figure BDA00034107817000000614
I.e. the output vector of the GRU network:
Figure BDA00034107817000000615
wherein the content of the first and second substances,
Figure BDA00034107817000000616
representing a sequence of actions of user u
Figure BDA00034107817000000617
The output vector after being processed by the GRU module,
Figure BDA00034107817000000618
representing the hidden state output by the hidden layer at the last moment of the GRU network; likewise, the output vector processed by the GRU module needs to be converted into the same dimensions as the representation vector:
Figure BDA00034107817000000619
wherein the content of the first and second substances,
Figure BDA00034107817000000620
representing a sequence of actions of user uThe vector is represented by a vector of values,
Figure BDA00034107817000000621
respectively representing a weight matrix and an offset for performing spatial transformation;
4-4 output layer: the output layer adds the results output by the heterogeneous knowledge map convolutional layer and the user behavior sequence sensing layer and then converts the results into output vectors with the same dimensionality as the predicted feature number;
Figure BDA0003410781700000071
o＝Wufinal+b
wherein u isfinalThe representation user finally represents the vector(s),
Figure BDA0003410781700000072
representing a representation vector with user neighbor characteristics learned by a heterogeneous knowledge graph convolutional layer,
Figure BDA0003410781700000073
the sequence of behaviors representing user u represents a vector,
Figure BDA0003410781700000074
an addition operation representing a vector; o denotes a user output vector, W denotes a weight matrix, and b denotes an offset vector;
and 5: performing softmax operation on the user output vector o obtained in the step 4 to obtain the probability corresponding to the basic information (namely the gender or the predicted age period) of the predicted user;
Figure BDA0003410781700000075
wherein, o'iRepresenting the probability representation of the i-th dimension obtained by the softmax function, oiA value representing the ith dimension of the output vector o; obtaining the user output vector corresponding to the user characteristics of the 0, 1, f-1 dimension through a softmax functionA probability representation;
the back propagation process of the whole model adopts a softmax cross entropy loss function, and the formula is as follows:
Figure BDA0003410781700000076
wherein the content of the first and second substances,
Figure BDA0003410781700000077
a set of users is represented as a set of users,
Figure BDA0003410781700000078
representing the cross-entropy loss function, yu and
Figure BDA0003410781700000079
respectively representing a real user tag value and a model predicted value;
Figure BDA00034107817000000710
is a regularization term of L2, λ represents a regularization coefficient for controlling the strength of the regularization of L2, and Θ represents parameters in the model, such as weight matrices between the user, entity, and relationship embedding matrices U, V, R, and the neural network layer.
The performance evaluation of the invention respectively adopts a MovieLens-1M movie data set and a Kyoto E-business data set. The model performs gender prediction two-classification performance evaluation and age prediction multi-classification performance evaluation on the two data sets respectively.
The following table shows the data volume of two data sets after the screening of the knowledge graph entities:
Figure BDA00034107817000000711
Figure BDA0003410781700000081
the two data sets respectively adopt Microsoft Satori and zhishi.me Chinese knowledge maps to conduct triple feature expansion on the entity set of the commodity name. The distribution of the user characteristics of each data set is as follows:
(1) sex aspect:
a) the ratio of male users to female users in the MovieLens-1M movie data set is 72 percent, and the ratio of male users to female users is 28 percent;
b) the data of the Jingdong E-business accounts for 44% of male users and 56% of female users.
(2) Age-related:
a) MovieLens-1M movie data set 22% of users under the age of 25, 35% of users between the age of 25 and 34, 29% of users between the age of 35 and 50, and 15% of users over 50;
b) in the data set of the Jingdong e-commerce, 14% of users under the age of 26, 55% of users under the age of 26 to 35, 30% of users under the age of 36 to 55, and 1% of users under the age of 55 are all users.
The performance evaluation indexes adopted by the invention are Accuracy and macro-F1.
True value 1 True value-1
Predicted value 1 TP(True Positive) FP(False Negative)
Prediction value-1 FN(False Negative) TN(True Negative)
Accuracy: the correctly classified samples account for the total number of samples:
Figure BDA0003410781700000082
macro _ F1 is a variant of the evaluation index F1_ score of a two-class model commonly used in machine learning, and the F1_ score evaluation index formula is as follows:
Figure BDA0003410781700000091
wherein precision and call respectively represent classification accuracy and recall, and respectively evaluate whether the classification of the model positive examples is accurate and the proportion of the positive examples judged by the classifier to all the positive examples, and from the above formula, it can be seen that F1_ score is an evaluation index combining the evaluation of the classifier accuracy and the recall.
Since the conventional F1_ score is mostly used for evaluating the second category, age prediction is a multi-category problem in experiments, macro _ F1 is used as an evaluation index, and macro _ F1 is an average value of each category F1_ score, namely:
Figure BDA0003410781700000092
wherein, F1_ score1，F1_score2，...，F1_scorenF1_ score for classes 1, 2,. N, respectively, N being the number of classes.
The following table shows the results of the gender prediction experiment of the present invention on the above two data sets:
Figure BDA0003410781700000093
the following table shows the results of the age prediction experiments of the present invention on the above two data sets:
Figure BDA0003410781700000094
Figure BDA0003410781700000101
in the above gender prediction and age prediction experimental result table, the logistic regression and support vector machine is a traditional machine learning classifier, the LightGBM is a gradient boosting decision tree-based efficient classification model proposed by microsoft, and the heterogeneous knowledge graph convolution network (Ba-KGCN) is a multi-source cross-border data fusion user portrait prediction model in the present invention.

Claims (9)

1. A user portrait prediction method based on multi-source cross-border data fusion is characterized by comprising the following steps:
step 1: collecting information generated by interaction of a user on a shopping platform, wherein the information comprises basic information of the user and user behavior records, and constructing a user set, a commodity name set and a commodity name-user interaction record;
basic information of the user comprises gender and age;
the user behavior records comprise time for purchasing commodities, commodity numbers and commodity names;
step 2: constructing a heterogeneous knowledge graph and a user historical interaction sequence;
2-1, constructing a heterogeneous knowledge graph and a user behavior sequence set
2-1-1, performing word segmentation on the commodity name to obtain a word segmentation result set { i }1，i2，...，im，...}，imRepresenting the mth participle;
2-1-2, performing 2-round recursion search on the segmentation result set in the public knowledge graph, discarding segmentation results which do not exist in the public knowledge graph, and forming an entity set epsilon (e) by using the remaining segmentation in the segmentation result set and entities searched in the public knowledge graph1，e2，...，en,., and further constructing them into triplets (i)m，contain，en) Cotain represents imAnd enThe incidence relation between the two; using the above triplet (i)m，contain，en) Constructing knowledge subgraphs corresponding to commodity names
Figure FDA0003410781690000011
2-1-3 knowledge subgraph corresponding to commodity name
Figure FDA0003410781690000012
Integrated into a heterogeneous knowledge graph
Figure FDA0003410781690000013
2-2, constructing a user-commodity name interaction matrix according to the user set, the commodity name set and the commodity name-user interaction records
Figure FDA0003410781690000014
N represents the number of users, and M represents the number of names of commodities;
2-3, further constructing a user history interaction sequence set according to the user-commodity name interaction matrix:
Figure FDA0003410781690000015
wherein
Figure FDA0003410781690000016
The name of the item representing the user u's historical interaction at the ith time,
Figure FDA0003410781690000017
represents user u and
Figure FDA0003410781690000018
the moment at which the interaction occurs;
and step 3: according to the user set and the entity set epsilonOne-step construction of user embedded matrix
Figure FDA0003410781690000019
Entity embedded matrix
Figure FDA00034107816900000110
Figure FDA00034107816900000111
And a user adjacency matrix
Figure FDA00034107816900000112
Wherein D represents a dimension of the vector; each element in the user adjacency matrix represents the similarity of the click behaviors of two users;
and 4, step 4: constructing a user portrait prediction model of multi-source cross-border data fusion;
the user portrait prediction model of multi-source cross-border data fusion comprises an input embedding layer, a heterogeneous knowledge graph convolution layer, a user behavior sequence perception layer and an output layer:
4-1 input embedding layer: constructing a set of user interaction entities N using a set of historical interaction sequences of userse(u); vectorizing and representing the user by using the user embedded matrix; vectorized representation S of user interaction entities using entity embedding matricesi(u); obtaining the user 'S adjacent user embedded vector according to the user' S adjacent matrix, and constructing the adjacent user set Su(u)；
4-2 heterogeneous knowledge map convolutional layer: after the representation vector of the user interaction entity enters the heterogeneous knowledge graph convolutional layer, two parts of operations are executed;
the 4-2-1 user interaction entity obtains a user-commodity name expression vector with neighbor features through H-round iterative aggregation of neighbor topological structure features
Figure FDA0003410781690000021
4-2-2 user u's neighboring user set expression vector and user u expression vector are aggregated to obtain user neighboringFeature representation vector
Figure FDA0003410781690000022
4-2-3 user-Commodity name representation vector
Figure FDA0003410781690000023
And user neighbor feature representation vector
Figure FDA0003410781690000024
Adding the spliced vector and the user u representation vector to obtain an output vector of the heterogeneous knowledge graph convolutional layer
Figure FDA0003410781690000025
4-3, the user behavior sequence perception layer adopts LSTM or GRU to model user sequence characteristics so as to extract potential interest of users; to be provided with
Figure FDA0003410781690000026
Obtaining a vector with the same dimensionality as the output of the heterogeneous knowledge map convolutional layer for input;
4-4 output layer: the output layer adds the results output by the heterogeneous knowledge map convolutional layer and the user behavior sequence sensing layer and then converts the results into output vectors with the same dimensionality as the predicted feature number;
and 5: and (4) performing softmax operation on the user output vector o obtained in the step (4) to obtain the probability corresponding to the basic information of the predicted user.
2. The method of claim 1, wherein the heterogeneous knowledge graph is used for user portrait prediction based on multi-source cross-border data fusion
Figure FDA0003410781690000027
Includes node V and edge E; the node V comprises a user set U, a commodity name set I and an entity set epsilon; the edges include three types, respectively entity-entity intellectual graph relationships
Figure FDA0003410781690000028
Commodity name-user interaction record EiuAnd the users have the same click behavior between pairs Euu(ii) a The entity-entity knowledge graph relationship
Figure FDA0003410781690000029
Is the relationship between any two entities in the entity set.
3. The method of claim 1, wherein the user portrait prediction layer is modeled by using user sequence features of LSTM:
hiding state of last moment of recurrent neural network
Figure FDA00034107816900000210
And cell status
Figure FDA00034107816900000211
Adding to obtain the output vector of the LSTM module:
Figure FDA00034107816900000212
wherein the content of the first and second substances,
Figure FDA00034107816900000213
representing historical interaction sequences of user u
Figure FDA00034107816900000214
The output vector after being processed by the LSTM module,
Figure FDA00034107816900000215
representing the cell state output by the LSTM neural network at the last moment,
Figure FDA00034107816900000216
representing the hidden state of the LSTM neural network output at the last time, T representing the last time,
Figure FDA00034107816900000217
represents an addition at the element level;
and then, carrying out spatial transformation on the output vector of the LSTM module, and converting the output vector into a user behavior sequence representation vector with the same dimension as the user representation vector:
Figure FDA00034107816900000218
wherein the content of the first and second substances,
Figure FDA00034107816900000219
the sequence of behaviors representing user u represents a vector,
Figure FDA00034107816900000220
and
Figure FDA00034107816900000221
respectively representing the weight matrix and the bias for spatial transformation, and P representing the number of LSTM hidden layer neurons.
4. The method of claim 1, wherein the user portrait prediction layer is modeled by using user sequence features of GRUs:
hidden state of last minute
Figure FDA0003410781690000031
I.e. the output vector of the GRU network:
Figure FDA0003410781690000032
wherein the content of the first and second substances,
Figure FDA0003410781690000033
representing a sequence of actions of user u
Figure FDA0003410781690000034
The output vector after being processed by the GRU module,
Figure FDA0003410781690000035
representing the hidden state output by the hidden layer at the last moment of the GRU network; likewise, the output vector processed by the GRU module needs to be converted into the same dimensions as the representation vector:
Figure FDA0003410781690000036
wherein the content of the first and second substances,
Figure FDA0003410781690000037
the sequence of behaviors representing user u represents a vector,
Figure FDA0003410781690000038
respectively representing the weight matrix and the offset for performing the spatial transformation.
5. The method of claim 1, wherein the output layer is specifically as follows:
Figure FDA0003410781690000039
o＝Wufinal+b
wherein u isfinalThe representation user finally represents the vector(s),
Figure FDA00034107816900000310
representing a representation vector with user neighbor characteristics learned by a heterogeneous knowledge graph convolutional layer,
Figure FDA00034107816900000311
the sequence of behaviors representing user u represents a vector,
Figure FDA00034107816900000312
an addition operation representing a vector; o denotes a user output vector, W denotes a weight matrix, and b denotes an offset vector.
6. The method for predicting the user portrait based on the multi-source cross-border data fusion as claimed in claim 1, wherein the step 5softmax operation is specifically:
Figure FDA00034107816900000313
wherein, o'iRepresenting the probability representation of the i-th dimension obtained by the softmax function, oiA value representing the ith dimension of the output vector o; and obtaining the probability representation of the user output vector pair corresponding to the user characteristics in the 0 th, 1 st and f-1 st dimensions through a softmax function.
7. The method of claim 1, wherein the backpropagation process of the multi-source cross-boundary data fusion-based user portrait prediction model adopts a softmax cross entropy loss function, and the formula is as follows:
Figure FDA00034107816900000314
wherein the content of the first and second substances,
Figure FDA0003410781690000041
a set of users is represented as a set of users,
Figure FDA0003410781690000042
representing the cross-entropy loss function, yu and
Figure FDA0003410781690000043
respectively representing a real user tag value and a model predicted value;
Figure FDA0003410781690000044
is the L2 regularization term, λ represents the regularization coefficient used to control the strength of the L2 regularization, and Θ represents the model parameters.
8. A multi-source cross-boundary data fusion-based user portrait prediction device, comprising a memory, a processor, and a multi-source cross-boundary data fusion-based user portrait prediction model program stored in the memory and executable on the processor, wherein when executed by the processor, the multi-source cross-boundary data fusion-based user portrait prediction model program implements the steps of any one of the above claims 1-7.
9. A storage medium storing a multi-source cross-boundary data fusion-based user portrait prediction model program, which when executed by a processor implements the steps of the multi-source cross-boundary data fusion-based user portrait prediction method of any one of claims 1 to 7.
CN202111531109.4A 2021-12-14 2021-12-14 User portrait prediction method based on multi-source cross-border data fusion Active CN114238758B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111531109.4A CN114238758B (en) 2021-12-14 2021-12-14 User portrait prediction method based on multi-source cross-border data fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111531109.4A CN114238758B (en) 2021-12-14 2021-12-14 User portrait prediction method based on multi-source cross-border data fusion

Publications (2)

Publication Number Publication Date
CN114238758A true CN114238758A (en) 2022-03-25
CN114238758B CN114238758B (en) 2023-04-11

Family

ID=80756046

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111531109.4A Active CN114238758B (en) 2021-12-14 2021-12-14 User portrait prediction method based on multi-source cross-border data fusion

Country Status (1)

Country Link
CN (1) CN114238758B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117829968A (en) * 2024-03-06 2024-04-05 南京数策信息科技有限公司 Service product recommendation method, device and system based on user data analysis

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112131404A (en) * 2020-09-19 2020-12-25 哈尔滨工程大学 Entity alignment method in four-risk one-gold domain knowledge graph
CN112350899A (en) * 2021-01-07 2021-02-09 南京信息工程大学 Network flow prediction method based on graph convolution network fusion multi-feature input
US20210256355A1 (en) * 2020-02-13 2021-08-19 International Business Machines Corporation Evolving graph convolutional networks for dynamic graphs
CN113590900A (en) * 2021-07-29 2021-11-02 南京工业大学 Sequence recommendation method fusing dynamic knowledge maps

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210256355A1 (en) * 2020-02-13 2021-08-19 International Business Machines Corporation Evolving graph convolutional networks for dynamic graphs
CN112131404A (en) * 2020-09-19 2020-12-25 哈尔滨工程大学 Entity alignment method in four-risk one-gold domain knowledge graph
CN112350899A (en) * 2021-01-07 2021-02-09 南京信息工程大学 Network flow prediction method based on graph convolution network fusion multi-feature input
CN113590900A (en) * 2021-07-29 2021-11-02 南京工业大学 Sequence recommendation method fusing dynamic knowledge maps

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
SIJIN ZHOU等: "Interactive Recommender System via Knowledge Graph-enhanced Reinforcement Learning", 《ARXIV》 *
刘欢: "基于知识图谱驱动的图神经网络推荐模型", 《计算机应用》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117829968A (en) * 2024-03-06 2024-04-05 南京数策信息科技有限公司 Service product recommendation method, device and system based on user data analysis
CN117829968B (en) * 2024-03-06 2024-05-31 南京数策信息科技有限公司 Service product recommendation method, device and system based on user data analysis

Also Published As

Publication number Publication date
CN114238758B (en) 2023-04-11

Similar Documents

Publication Publication Date Title
Ma et al. Machine learning and AI in marketing–Connecting computing power to human insights
Khan et al. CNN with depthwise separable convolutions and combined kernels for rating prediction
US20220301024A1 (en) Sequential recommendation method based on long-term and short-term interests
CN108648049B (en) Sequence recommendation method based on user behavior difference modeling
Cao Coupling learning of complex interactions
Pan et al. Study on convolutional neural network and its application in data mining and sales forecasting for E-commerce
CN111222332B (en) Commodity recommendation method combining attention network and user emotion
CN110674407B (en) Hybrid recommendation method based on graph convolution neural network
Murphy Machine learning: a probabilistic perspective
CN107833117B (en) Bayesian personalized sorting recommendation method considering tag information
CN106447066A (en) Big data feature extraction method and device
CN112487199B (en) User characteristic prediction method based on user purchasing behavior
CN106445988A (en) Intelligent big data processing method and system
Hossein Javaheri Response modeling in direct marketing: a data mining based approach for target selection
CN108921602B (en) User purchasing behavior prediction method based on integrated neural network
CN114238758B (en) User portrait prediction method based on multi-source cross-border data fusion
Meng et al. Concept-concept association information integration and multi-model collaboration for multimedia semantic concept detection
Zhu et al. Multimodal sparse linear integration for content-based item recommendation
Zhang et al. Image annotation of ancient Chinese architecture based on visual attention mechanism and GCN
CN116823321B (en) Method and system for analyzing economic management data of electric business
Ifada et al. Do-rank: DCG optimization for learning-to-rank in tag-based item recommendation systems
CN115293851A (en) Recommendation method for introducing item category information into graph neural network
DIVYA et al. Matrix factorization for movie recommended system using deep learning
Svensson et al. Exploring NMF and LDA Topic Models of Swedish News Articles
Tekchandani et al. Applied Deep Learning: Design and implement your own Neural Networks to solve real-world problems (English Edition)

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