CN113343094A - Information-enhanced meta-learning method for relieving cold start problem of recommended user - Google Patents

Information-enhanced meta-learning method for relieving cold start problem of recommended user Download PDF

Info

Publication number
CN113343094A
CN113343094A CN202110683807.XA CN202110683807A CN113343094A CN 113343094 A CN113343094 A CN 113343094A CN 202110683807 A CN202110683807 A CN 202110683807A CN 113343094 A CN113343094 A CN 113343094A
Authority
CN
China
Prior art keywords
user
meta
information
cold start
item
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110683807.XA
Other languages
Chinese (zh)
Inventor
吴国栋
毕海娇
汪菁瑶
涂立静
李景霞
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.)
Anhui Agricultural University AHAU
Original Assignee
Anhui Agricultural University AHAU
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 Anhui Agricultural University AHAU filed Critical Anhui Agricultural University AHAU
Priority to CN202110683807.XA priority Critical patent/CN113343094A/en
Publication of CN113343094A publication Critical patent/CN113343094A/en
Pending legal-status Critical Current

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
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/284—Lexical analysis, e.g. tokenisation or collocates
    • 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
    • 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

Abstract

The invention discloses a meta-learning method for relieving the problem of recommending user cold start by information enhancement, which comprises the steps of constructing a user-object bipartite graph according to the comment relationship of user objects; taking user neighborhood information and an article description text as sources of node information, randomly sampling the constructed bipartite graph, constructing meta-learning tasks, and simulating a scene for realizing recommendation for a new user by each task representing cold start recommendation for the new user; then, a bert method is used for extracting the characteristics of the text data aiming at each meta-learning task so as to obtain preference information; and guiding the meta global parameters to generate local parameters of the embedded generation function of each user according to the preference information, inputting the meta learning task into the recommendation model to obtain the prediction scores of the user on the articles, updating the meta learning parameters, and directly using the trained parameters to the untrained new user. The method and the device for recommending the preference relieve the problem that the preference recommendation cannot be evaluated accurately due to few new user interactions.

Description

Information-enhanced meta-learning method for relieving cold start problem of recommended user
Technical Field
The invention relates to the field of network information pushing methods, in particular to a meta-learning method for relieving the problem of cold start of a recommended user by enhancing information.
Background
With the development of the internet, the e-commerce platform can almost meet any shopping requirements of people, and more people rely on the online shopping. In the face of explosively increasing numbers of users and items, recommendation systems have come up. The recommendation system provides accurate personalized recommendation for the user according to the preference of the user, and improves the shopping experience of the user and the sales volume of the e-commerce platform. But recommendation models are generally sensitive to user-specific data, perform better for some users, but generally have poor generalization capabilities. And the recommendation model needs a large amount of user interaction data to train to provide personalized recommendation, and the biggest characteristic of a new user is that the user interaction data is less, so that the traditional method has a serious cold start problem. The cold start problem and the poor generalization capability problem of the recommendation system need to be solved urgently.
The latest research applies an emerging meta-learning algorithm to recommendation, mainly simulates new users with few interactions to relieve the problem in the cold start aspect by a method for realizing small sample learning through meta-learning, and is also applied to aspects such as adaptive selection or retraining of the existing recommendation system. At present, most of researches on the problem of relieving cold start of Meta-Learning directly utilize the idea of Model-independent Meta-Learning algorithm (MAML) to carry out Learning on parameter initialization of a recommended Model, training is carried out on a large number of different tasks, and the representation adaptive to the new task can be quickly obtained through a small number of gradient steps. However, the existing meta-learning research on the cold start problem of the recommendation system has some problems, the complete cold start scene cannot be well relieved, most of the research is based on the fact that a new user has a small amount of interaction, and the meta-learning method has the defects of poor stability, local optimal solution obtained by a model on some users and the like. Therefore, some user preference information is required to be added to guide the parameters of meta-learning to guide generation of recommended model parameters of different users, so that the generalization capability of the model is improved. But for privacy protection reasons the user's profile has difficulty obtaining even or getting false profile information, resulting in difficulties in initial preference acquisition.
Disclosure of Invention
The invention aims to provide an information-enhanced meta-learning method for relieving the problem of cold start of recommended users, which is optimized by extracting preference information from comment texts of users on articles, so that the problem of cold start of users in a recommendation system in the prior art is solved, and more accurate personalized recommendation is provided for new users.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
an information-enhanced meta-learning method for alleviating the problem of cold start of a recommended user comprises the following steps:
(1) constructing an adjacency matrix of the user and the article based on the comment relationship, and then constructing a user-article interaction bipartite graph based on the adjacency matrix, wherein the user and the article are respectively used as nodes in the user-article interaction bipartite graph;
(2) generating an embedded vector of an article node according to the attribute characteristics of the article, and generating an embedded vector of a user node by using an aggregation method;
(3) sampling the user-item interaction bipartite graph to construct a data set of a meta-learning task;
(4) extracting user preference information from comment texts of users on articles by using a bert model;
(5) building a full-connection network as a recommendation model, adding preference information of the user nodes into the recommendation model, training the recommendation model by adopting the data set obtained in the step (3), and updating global parameters of the recommendation model to optimal parameters based on training results;
(6) and recommending the articles for the new user by using the recommendation model with the global parameter as the optimal parameter.
Further, in the step (1), a comment text of the user on the article is acquired, and a comment relation between the user and the article is obtained based on the comment text.
Further, in the step (1), the adjacency matrix is used as an edge of the user-item interaction bipartite graph.
Further, in the step (2), an embedding vector of the item node is generated according to the attribute feature of the item itself.
Further, in the step (2), information contained in a neighbor item node of each user in the user-item interaction bipartite graph is aggregated by using an aggregation function to generate an embedded vector of the user node.
Further, in the step (3), the data set of the meta-learning task includes a support set and a query set, wherein:
selecting part of user nodes in the user-article interaction bipartite graph, and randomly sampling a plurality of first-order neighbors for each selected user respectively, namely selecting the first-order neighbors as history interaction articles for each user node, thereby forming a training data set, namely a support set, of a meta-training task;
in the process of sampling the support set for each user, the neighbors which are not sampled by the user to the first order are sampled to form a test data set, namely a query set, of the meta-training task, so that the support set and the query set are ensured to be mutually independent, and the missnooping of data is prevented.
Further, in the bert model in the step (4), word embedding of comment texts is generated for a plurality of historical interactive articles of each user, then the word embedding is trained by using a language model, that is, feature extraction is performed on the word embedding by using a bidirectional Transformer in the bert model to obtain new vectors, and after the new vectors obtained by each user are spliced, user preference information is obtained.
Further, in the step (6), a recommendation model with global parameters as optimal parameters is used to calculate a prediction score between each user and each article, and article recommendation is performed on new users according to the prediction scores.
Compared with the prior art, the invention has the advantages that:
the method extracts the characteristic information of each user comment text, extracts partial preference information of each specific user, increases the identifiability of the meta global parameter when initializing different users, and improves the stability and generalization capability of a recommendation model;
2. according to the method, the number of the first-order neighbor nodes of the sub-graph sampling is controlled, new users are simulated, and the global parameters of the recommendation system, which have high generalization capability for most of the new users, are learned, so that the recommendation system is quickly adapted to the new users to relieve the recommendation cold start problem;
3. the method has great competitiveness on the warm start problem, particularly when recommendation is realized on a data set with high sparsity, the warm start effect is superior to that of a part of traditional recommendation methods, and the warm start effect is better along with the sparsity of the data set.
Drawings
FIG. 1 is a bipartite view of a user and an article in accordance with the present invention;
FIG. 2 is a schematic diagram of bert extracting preference information in a user comment text in the present invention;
fig. 3 is a schematic diagram of the present invention.
Detailed Description
As shown in fig. 3, the meta-learning method for alleviating the recommended user cold-start problem with information enhancement according to the embodiment is performed according to the following steps:
step 1, obtaining a comment text of a user on an article, taking the behavior of the user on the article as a basis for the contact between the user and the article, wherein the basis is referred to as a comment relation between the user and the article for short, and constructing a user-article interaction bipartite graph according to all the users, the article and the comment relation thereof as shown in fig. 1, wherein the specific process is as follows:
step 1.1, according to the user set U of all users, U ═ U1,u2,...,uq,...,umIn which uqE.u, q 1,21,i2,...,iv,...,inIn which ivE, I, v is 1,2, n represents different articles, an adjacency matrix of the user and the articles is constructed based on the comment relation, and if the user comments the articles, the adjacency matrix R corresponds to the position RmnA value of 1, in contrast to rmnThe value is 0. Therefore, an adjacency matrix M of the user goods based on the comment relation is obtainedUI＝{rmn}，
Figure BDA0003123555320000041
du、diRespectively corresponding to the vector dimensions representing the user and the article,
Figure BDA0003123555320000042
is the dimension of the adjacency matrix.
Step 1.2, according to the adjacency matrix M of users and articlesUIBuilding a user-item interaction bipartite graph G ═ { UU ^ I ^ MUIWherein U and I are user set and item set respectively, and constitute user node and item node in the graph respectively, MUIRepresenting edges in the graph for the interaction between each user item;
step 2, obtaining the embedded vector of each user and each article node, wherein the specific process is as follows:
step 2.1, generating embedded vector representation e of item nodes according to attribute characteristics of each item by using configuration files with characteristic information of each itemi；
Step 2.2, sampling the neighbor item nodes of the quantity j of each user u by using an aggregation function
Figure BDA0003123555320000043
Aggregating the information contained in the neighbor object nodes to obtain the embedded vector representation e of the user nodeu：
Figure BDA0003123555320000044
In the formula (1), huIs a user node one-hot vector representation,
Figure BDA0003123555320000045
is an embedded representation set of sampled neighboring item nodes, where
Figure BDA0003123555320000046
Is an initial embedded representation, i, of each item sampleda∈I,a＝1,2,...,j,iaIs an item in defined item set I;
step 3, sampling the user-article bipartite graph and constructing a meta-learning task
Figure BDA0003123555320000047
Data set D ofuThe specific process is as follows:
step 3.1, selecting part of user nodes, randomly sampling k first-order neighbor item nodes for each selected user node u, and generating a subgraph G of each user nodeu，G→GuI.e. selecting k historical interactive articles for each user node to form a meta-training task
Figure BDA0003123555320000048
Is a support set
Figure BDA0003123555320000049
Step 3.2, in the process of sampling the support set for each user, sampling the neighbors which are not sampled by the user and have first order to form a test data set, namely a query set, of the meta-training task
Figure BDA00031235553200000410
The mutual independence of the support set and the query set is ensured, and the data snooping error is prevented.
Thus, meta-learning task
Figure BDA0003123555320000051
Of the data set
Figure BDA0003123555320000052
Step 4, as shown in fig. 2, extracting user preference information in each meta-learning task simulating cold start recommendation for a new user by using a bert model, wherein the specific process is as follows:
step 4.1, generating word embedding T of comment text for k historical interactive articles selected by user nodes in each meta-learning tasku＝{t1,t2,...,tx,...,tkWhere t isxA comment text word embedded representation representing each item, x ═ 1, 2.., k;
step 4.2, training word embedding by using the language model, namely extracting features of word embedding by using a bidirectional Transformer in a bert model to obtain a new vector Su＝{s1,s2,...,sz,...,skIn which s isz∈SuEmbedding words representing the comment text of each article into a new vector obtained by bidirectional transform processing, wherein z is 1,2uGenerating preference information of the user node;
and 5, building a full-connection network as a recommendation model, adding the preference information of the user node extracted by the bert algorithm for guidance, training and updating parameters, wherein the specific process is as follows:
and 5.1, constructing a full connection layer as a user embedded generator, and randomly initializing a meta-learning global parameter phi which is { theta, omega }, wherein the meta-global parameter phi is composed of two parts of theta and omega, theta represents all trainable parameter combinations of the user embedded generator, and omega represents all trainable parameter combinations of the recommended model. Preference information B of user node extracted by bert algorithmuGlobal parameter theta for guiding users to embed into generator generates local parameter theta for each user to integrate into personalized informationuUpdating the initial embedding e of the user by personalized local parametersuAs shown in the following formula:
θu←θ+αBu (2)，
Figure BDA0003123555320000053
in the formula (2), alpha is a hyper-parameter and controls the influence of the extracted preference information on the personalized parameters; in the formula (3)
Figure BDA0003123555320000054
Is given by the parameter θuFull connection layer network of user embedded generator, xuIs prepared by
Figure BDA0003123555320000055
Updated user embedding;
step 5.2, generating a personalized local parameter omega of the recommendation model for each user through a global parameter phi ═ { theta, omega } of the recommendation modelu←ω；
Step 5.3, embedding x into each useruAnd corresponding historical interactive item embedding eiI 1, 2.., k, input to the recommendation model, calculate a prediction score
Figure BDA0003123555320000061
Calculating the corresponding loss of each user through the prediction score and the real score y represented by the label
Figure BDA0003123555320000062
As shown in the following equation:
Figure BDA0003123555320000063
in formula (4)
Figure BDA0003123555320000064
Is a parameter of ωuThe recommendation model of (1), namely the fully connected layer recommended by the user article;
step 5.4, reversely updating the local parameter theta through lossu、ωuAs shown in the following formula:
Figure BDA0003123555320000065
Figure BDA0003123555320000066
both beta and gamma in the expressions (5) and (6) are hyper-parameters and are used for controlling the learning rate during parameter updating; formulas (5) and (6) in-meta learning task
Figure BDA0003123555320000067
Support set of
Figure BDA0003123555320000068
Upper calculated loss to pair thetau、ωuCarrying out local updating;
and 5.5, after all the user nodes are trained, updating the global parameter phi by using loss and reverse direction, wherein the global parameter phi is represented by the following formula:
Figure BDA0003123555320000069
equation (7) set of queries in a task
Figure BDA00031235553200000610
Calculating the loss and updating the parameter phi globally, wherein eta is a hyper-parameter;
and 6, acquiring an embedded vector representation of new user fusion preference information by using a global parameter theta learned by a user embedded generator part through training, initializing a recommendation model parameter by using the global parameter omega learned by a recommendation model part, inputting the acquired user embedding and an article to be recommended for the user into the recommendation model, calculating a prediction score between each user and the article by using the model, and generating a recommended article list by using a Top-n sequencing mode according to the acquired prediction score to realize recommendation.
The embodiments of the present invention are described only for the preferred embodiments of the present invention, and not for the limitation of the concept and scope of the present invention, and various modifications and improvements made to the technical solution of the present invention by those skilled in the art without departing from the design concept of the present invention shall fall into the protection scope of the present invention, and the technical content of the present invention which is claimed is fully set forth in the claims.

Claims (8)

1. An information-enhanced meta-learning method for alleviating the problem of cold start of a recommended user, comprising the steps of:
(1) constructing an adjacency matrix of the user and the article based on the comment relationship, and then constructing a user-article interaction bipartite graph based on the adjacency matrix, wherein the user and the article are respectively used as nodes in the user-article interaction bipartite graph;
(2) generating an embedded vector of an article node according to the attribute characteristics of the article, and generating an embedded vector of a user node by using an aggregation method;
(3) sampling the user-item interaction bipartite graph to construct a data set of a meta-learning task;
(4) extracting user preference information from comment texts of users on articles by using a bert model;
(5) building a full-connection network as a recommendation model, adding preference information of the user nodes into the recommendation model, training the recommendation model by adopting the data set obtained in the step (3), and updating global parameters of the recommendation model to optimal parameters based on training results;
(6) and recommending the articles for the new user by using the recommendation model with the global parameter as the optimal parameter.
2. The meta-learning method for relieving the recommended user cold start problem with enhanced information according to claim 1, wherein in the step (1), a comment text of the user on the item is obtained, and a comment relationship between the user and the item is obtained based on the comment text.
3. The meta-learning method for relieving the recommended user cold start problem with enhanced information according to claim 1, wherein in the step (1), the adjacency matrix is used as the edge of the user-item interaction bipartite graph.
4. The meta-learning method for relieving the cold start problem of the recommended users with enhanced information as claimed in claim 1, wherein in the step (2), the embedded vector of the item node is generated according to the attribute feature of the item itself.
5. The meta-learning method for relieving the recommended user cold start problem with information enhancement as claimed in claim 1, wherein in the step (2), the information contained in the neighbor item node of each user in the user-item interaction bipartite graph is aggregated by using an aggregation function to generate the embedded vector of the user node.
6. The meta-learning method for alleviating the recommended user cold start problem with information enhancement as claimed in claim 1, wherein in the step (3), the data set of the meta-learning task comprises a support set and a query set, which respectively correspond to a training set and a test set in a data set of a conventional machine learning task, wherein:
selecting part of user nodes in the user-article interaction bipartite graph, and randomly sampling a plurality of first-order neighbors for each selected user respectively, namely selecting the first-order neighbors as history interaction articles for each user node, thereby forming a training data set, namely a support set, of a meta-training task;
in the process of sampling the support set for each user, the neighbors which are not sampled by the user to the first order are sampled to form a test data set, namely a query set, of the meta-training task, so that the support set and the query set are ensured to be mutually independent, and the missnooping of data is prevented.
7. The meta-learning method for relieving the cold start problem of recommended users with enhanced information as claimed in claim 1 or 6, wherein in the bert model of step (4), word embedding of comment texts is first generated for a plurality of historical interactive articles of each user, and then the word embedding is trained by using a language model, that is, feature extraction is performed on the word embedding by using a bidirectional Transformer in the bert model to obtain new vectors, and after the new vectors obtained by each user are spliced, user preference information is obtained.
8. The meta-learning method for relieving the cold start problem of recommended users with enhanced information as claimed in claim 1, wherein in the step (6), a recommendation model with global parameters as optimal parameters is used to calculate a prediction score between each user and an item, and the item is recommended to a new user according to the prediction score.
CN202110683807.XA 2021-06-21 2021-06-21 Information-enhanced meta-learning method for relieving cold start problem of recommended user Pending CN113343094A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110683807.XA CN113343094A (en) 2021-06-21 2021-06-21 Information-enhanced meta-learning method for relieving cold start problem of recommended user

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110683807.XA CN113343094A (en) 2021-06-21 2021-06-21 Information-enhanced meta-learning method for relieving cold start problem of recommended user

Publications (1)

Publication Number Publication Date
CN113343094A true CN113343094A (en) 2021-09-03

Family

ID=77477974

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110683807.XA Pending CN113343094A (en) 2021-06-21 2021-06-21 Information-enhanced meta-learning method for relieving cold start problem of recommended user

Country Status (1)

Country Link
CN (1) CN113343094A (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140176565A1 (en) * 2011-02-17 2014-06-26 Metail Limited Computer implemented methods and systems for generating virtual body models for garment fit visualisation
CN111931045A (en) * 2020-07-30 2020-11-13 北京邮电大学 Heterogeneous information network cold start recommendation method and device based on meta-learning
CN112650929A (en) * 2020-12-31 2021-04-13 安徽农业大学 Graph neural network recommendation method integrating comment information

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140176565A1 (en) * 2011-02-17 2014-06-26 Metail Limited Computer implemented methods and systems for generating virtual body models for garment fit visualisation
CN111931045A (en) * 2020-07-30 2020-11-13 北京邮电大学 Heterogeneous information network cold start recommendation method and device based on meta-learning
CN112650929A (en) * 2020-12-31 2021-04-13 安徽农业大学 Graph neural network recommendation method integrating comment information

Similar Documents

Publication Publication Date Title
Luce Artificial intelligence for fashion: How AI is revolutionizing the fashion industry
Pei et al. Interacting attention-gated recurrent networks for recommendation
Truong et al. Multimodal review generation for recommender systems
CN103164463B (en) Method and device for recommending labels
CN105045907B (en) A kind of construction method of vision attention tagging user interest tree for Personalized society image recommendation
CN109657204B (en) Auto-pairing fonts using asymmetric metric learning
CN107862561A (en) A kind of method and apparatus that user-interest library is established based on picture attribute extraction
CN109544306A (en) A kind of cross-cutting recommended method and device based on user behavior sequence signature
CN107169031B (en) Picture material recommendation method based on depth expression
CN104199818A (en) Social recommending method based on classification
CN112966091A (en) Knowledge graph recommendation system fusing entity information and heat
CN107766439A (en) A kind of personalized recommendation method of fusion structure feature and implicit feedback
Jiang et al. Visual font pairing
Xia et al. A regularized optimization framework for tag completion and image retrieval
KR102301663B1 (en) Identifying physical objects using visual search query
Xiong et al. Affective impression: Sentiment-awareness POI suggestion via embedding in heterogeneous LBSNs
Wang et al. Fg-t2m: Fine-grained text-driven human motion generation via diffusion model
Li et al. Graph-based relation-aware representation learning for clothing matching
CN109858031A (en) Neural network model training, context-prediction method and device
CN112052388A (en) Method and system for recommending gourmet stores
CN113343094A (en) Information-enhanced meta-learning method for relieving cold start problem of recommended user
CN110020195A (en) Article recommended method and device, storage medium, electronic equipment
CN114238752A (en) Article recommendation method and device and storage medium
Trivedi Machine Learning Fundamental Concepts
El-Gayar Automatic Generation of Image Caption Based on Semantic Relation using Deep Visual Attention Prediction

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210903