CN115422442A - Cold-start recommendation-oriented anti-self-coding migration learning method - Google Patents

Cold-start recommendation-oriented anti-self-coding migration learning method Download PDF

Info

Publication number
CN115422442A
CN115422442A CN202210976839.3A CN202210976839A CN115422442A CN 115422442 A CN115422442 A CN 115422442A CN 202210976839 A CN202210976839 A CN 202210976839A CN 115422442 A CN115422442 A CN 115422442A
Authority
CN
China
Prior art keywords
user
positive
hypergraph
warm
negative
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
CN202210976839.3A
Other languages
Chinese (zh)
Other versions
CN115422442B (en
Inventor
吴汉瑞
龙锦益
李诺思
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.)
Jinan University
Original Assignee
Jinan 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 Jinan University filed Critical Jinan University
Priority to CN202210976839.3A priority Critical patent/CN115422442B/en
Publication of CN115422442A publication Critical patent/CN115422442A/en
Application granted granted Critical
Publication of CN115422442B publication Critical patent/CN115422442B/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
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2457—Query processing with adaptation to user needs
    • G06F16/24578—Query processing with adaptation to user needs using ranking
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Computational Linguistics (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (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 discloses a cold start recommendation-oriented anti-self-coding transfer learning method which is used for processing the problem of article cold start through five steps and firstly constructing a positive and negative hypergraph. And then acquiring the conventional characteristics of warm users and cold users by utilizing a multilayer perceptron. And constructing a hypergraph self-encoder, respectively acquiring positive and negative characteristics of the warm user and the article, and reconstructing the positive and negative hypergraphs. A matching discriminator is developed to minimize the loss of classification of the positive and negative features of the warm user and the distribution gap between the positive and regular features of the warm user. Therefore, the positive characteristics of the warm user and the normal characteristics of the goods are connected, the positive characteristics of the warm user and the positive characteristics of the goods are related through a positive hypergraph, the relation between the normal characteristics of the cold start user and the positive characteristics of the goods is found, euclidean distances of the normal characteristics of the cold user and the positive characteristics of the goods are calculated and ranked, top-K goods are recommended to the user, and the superiority of the method is verified on a plurality of real data sets by using the indexes of precision, recall ratio, NDCG and hit ratio. The method effectively solves the cold start problem and provides more accurate and personalized recommendation suggestions for the user.

Description

Cold-start recommendation-oriented anti-self-coding migration learning method
Technical Field
The invention relates to the technical field of information recommendation, in particular to a cold start recommendation oriented anti-self-coding transfer learning method.
Background
With the continuous development of the internet, people enjoy the convenience of computer technology and simultaneously generate and manufacture massive data. The problem of information overload caused by these large-scale and complex data increases the difficulty for users to capture useful information in tasks such as information retrieval and electronic commerce. The recommendation system, as an information filtering tool, plays a crucial role in providing accurate and personalized recommendations for users, and has been receiving wide attention in the past decade.
The traditional recommendation method based on collaborative filtering greatly depends on historical interaction data of users and articles, and effective recommendation suggestions cannot be provided for new users or new articles, namely the cold start problem exists. Researchers have proposed some methods of deep learning to address the cold start problem by discovering potential relationships between users and items. Recently, hypergraph learning has received a wide attention due to excellent modeling ability on complex data, and researchers have introduced hypergraph learning into recommendation systems. In the hypergraph, one hyperedge can connect a plurality of nodes; in the user interaction information of the recommendation system, one article can interact with a plurality of users. Therefore, the hypergraph can naturally express the high-level complex relation between the user and the article, namely the value of 1 in the hypergraph represents interaction, and the value of 0 represents no interaction, so that the high-level feature representation of the user and the article can be learned by utilizing the hypergraph convolution network.
Domain adaptation aims to leverage knowledge from other source domains to assist the learning task of the target domain, and the main challenge of domain adaptation is to reduce the domain gap between the source domain and the target domain. Existing work tends to employ methods such as maximum average difference, optimal transmission, and combat loss to estimate the distribution difference between two domains. But not accurate and personalized in dealing with the user cold start and item cold start recommendation problems.
Therefore, it is necessary to invent a method of counteracting the self-encoding migration learning for cold start recommendation to solve the above problems.
Disclosure of Invention
The invention aims to provide a cold-start recommendation-oriented anti-self-coding migration learning method, and aims to provide more accurate and personalized recommendation suggestions for users, so as to solve the problems of cold start of the users and cold start of articles.
In order to achieve the above purpose, the invention provides the following technical scheme: a cold-start recommendation-oriented anti-self-coding transfer learning method comprises the following steps:
s1, constructing a hypergraph, modeling interactive information of a user and an article, and designing a positive hypergraph and a negative hypergraph according to the original hypergraph;
s2, designing a multi-layer perceptron network to obtain conventional feature representations of warm users and cold start users;
and S3, constructing a hypergraph self-encoder for respectively using the positive hypergraph and the negative hypergraph, acquiring positive and negative feature representations of the warm user and the warm object, and regarding the positive features as source data and the conventional features as target data. In addition, positive and negative hypergraphs are reconstructed and used for storing the associated information between the user and the article;
s4, constructing a matching discriminator, distributing pseudo labels for the positive and negative characteristics of the warm user, and minimizing the classification loss of the positive and negative characteristics of the warm user and the distribution difference between the positive characteristics and the conventional characteristics;
s5, calculating Euclidean distances between the conventional features and the article features of the cold-start user, sequencing, recommending Top-K articles to the user, and calculating precision, recall rate, NDCG and hit rate;
the step S1 includes:
s1.1, preprocessing a user scoring matrix, deleting articles with the score of 0, and reserving the rest articles as interactive items;
s1.2, enabling a user to perform the following steps according to the following steps of 9:1 into warm users and cold start users, and constructing a hypergraph according to the interactive information of the warm users and the articles, specifically
Figure BDA0003798787200000021
And the form of the hypergraph is as follows:
Figure BDA0003798787200000031
wherein n is w Representing the number of warm users, m represents the total number of items, the value of R (i, j) is 1 representing that the warm user i has interaction with the item j, otherwise the value of R (i, j) is 0;
s1.3, positive superscript R + Equal to the original hypergraph R, the positive hypergraph represents the preference items of the user, and the items with the value of R (i, j) of 0 are randomly selected from the original hypergraph R to construct a negative hypergraph R - ；
The form of the negative hypergraph is then as follows:
Figure BDA0003798787200000032
the step S2 includes:
s2.1, constructing a multilayer perceptron neural network to obtain the conventional characteristics of the warm user and the cold start user, wherein the specific formula is as follows:
Figure BDA0003798787200000033
where h (-) denotes a fully connected neural network,
Figure BDA0003798787200000034
representing the characteristics of users, the trust relations of the users exist in the original data, the trust relations are used as the trust relations, n represents the total number of the users, d represents the dimension of the characteristics, phi represents the trainable network weight and has
U＝[U w ,U c ]
Figure BDA0003798787200000035
U w And U c Representing the original characteristics of a warm user and a cold start user respectively,
Figure BDA0003798787200000036
and
Figure BDA0003798787200000037
conventional characterizations representing warm and cold start users, respectively.
The step S3 includes:
s3.1, the relation between the user and the article is represented by a hypergraph, and the hypergraph self-encoder is used for learning high-level feature representation of the warm user and the article on the hypergraph, wherein the specific formula is as follows:
Figure BDA0003798787200000038
wherein f is + Is a hypergraph self-encoder, T is the characteristic of the article,
Figure BDA0003798787200000039
in order to reconstruct the hypergraph,
Figure BDA00037987872000000310
is a positive feature of warming the user,
Figure BDA0003798787200000041
is a positive feature of the article;
s3.2, reconstructing a positive hypergraph by using the positive characteristics of the warm user and the characteristics of the article acquired in the step S3.1, wherein the loss of the reconstructed positive hypergraph is represented as:
Figure BDA0003798787200000042
wherein
Figure BDA0003798787200000043
Is a binary cross entropy function.
S3.3, learning high-level feature representation of warm users and articles by using a hypergraph self-encoder on the negative hypergraph, wherein a specific formula is as follows:
Figure BDA0003798787200000044
wherein f is - Is a hypergraph self-encoder.
Figure BDA0003798787200000045
In order to reconstruct the hypergraph,
Figure BDA0003798787200000046
is a negative feature of a warm user,
Figure BDA0003798787200000047
is a negative characteristic of the article;
s3.4, reconstructing a negative hypergraph by using the negative characteristics of the warm user and the characteristics of the object acquired in the step S3.3, wherein the loss of the reconstructed negative hypergraph can be expressed as follows:
Figure BDA0003798787200000048
the step S4 includes:
s4.1 Positive characteristics for warming users respectively
Figure BDA0003798787200000049
And negative characteristic
Figure BDA00037987872000000410
Assigning pseudo-labels, i.e. setting
Figure BDA00037987872000000411
Assign a pseudo label as
Figure BDA00037987872000000412
Wherein n is w The number of warm users, d is the characteristic dimension;
the positive characteristics comprise preference information of the user, and the negative characteristics comprise information which is not interested by the user;
s4.2, positive feature to bridge Warm user and regular feature, setting
Figure BDA00037987872000000413
Assign a pseudo label as
Figure BDA00037987872000000414
S4.3, designing a matching discriminator, and minimizing the classification loss of the positive and negative characteristics of the warm user and the distribution difference between the positive characteristics and the conventional characteristics according to the following formula:
Figure BDA0003798787200000051
where a is a trainable parameter which is,
Figure BDA0003798787200000052
and
Figure BDA0003798787200000053
each represents
Figure BDA0003798787200000054
And
Figure BDA0003798787200000055
the (j) th row vector of (c),
Figure BDA0003798787200000056
and
Figure BDA0003798787200000057
respectively represent
Figure BDA0003798787200000058
And
Figure BDA0003798787200000059
value of j row of (1), G y (. Is a classifier, G) c (ii) a characteristic matcher,
Figure BDA00037987872000000510
is a loss function of the classifier and,
Figure BDA00037987872000000511
representing the penalty function between domains;
s4.4, by step S4.2, the positive and negative characteristics of the warm user are separated, the positive characteristics are connected with the regular characteristics, thereby bridging the regular characteristics of the cold start user and the item characteristic image recommendation, introducing a gradient back propagation layer, following the following formula:
Figure BDA00037987872000000512
s4.5, integrating the steps, designing an overall loss function of the model for training, wherein the formula is as follows:
Figure BDA00037987872000000513
where β and η are adjustable parameters to balance the weights between the three loss functions;
s4.6, repeating the steps S4.3 and S4.4 until convergence, and connecting the positive characteristic and the conventional characteristic of the warm user;
the positive characteristics of the warm user and the positive characteristics of the article are connected through a positive hypergraph, the normal characteristics of the warm user and the normal characteristics of the cold start user have similar distribution, and the normal characteristics of the cold start user and the positive characteristics of the article are observed and recorded.
The step S5 includes:
s5.1, discovering the relation between the cold start user and the articles, calculating the Euclidean distance between the conventional characteristics and the article characteristics of the cold start user, recommending Top-K articles to the user, and evaluating the performance of the method by adopting the following four indexes:
A. precision: the ratio of the total number of the items in the recommendation list is calculated, all users are averaged, and the larger the value is, the better the value is, the calculation formula is as follows:
Figure BDA00037987872000000514
wherein N is ts Is the number of cold start users, L i Is the ith cold start user's true favorite item,
Figure BDA0003798787200000061
is the recommended Top-K items;
B. the recall ratio is as follows: as the proportion of correctly recommended articles to the total number of articles to be recommended, the larger the value is, the better the value is, and the calculation formula is as follows:
Figure BDA0003798787200000062
c, NDCG: for measuring the superiority of the recommendation list, when the result with high relevance appears at a more front position, the higher the index is, the calculation formula is as follows:
Figure BDA0003798787200000063
wherein r is i Is the correlation of the ith item, the user prefers the item i then r i The value is 1, otherwise 0, and the value of IDCG ensures that the value of NDCG can be atBetween 0 and 1;
D. the hit rate: as a commonly used index for measuring the recall ratio, the larger the value is, the better the value is, and the calculation formula is as follows:
Figure BDA0003798787200000064
in the technical scheme, the invention provides the following technical effects and advantages:
1. the negative hypergraph is introduced to learn the negative high-dimensional characteristic representation so as to assist in acquiring the high-dimensional characteristic representation of the user and the object, and the proposed model can effectively solve the cold start problem;
2. the invention applies the hypergraph automatic encoder to the positive hypergraph and the negative hypergraph to respectively obtain positive and negative feature representations, and also develops a matching discriminator to minimize the classification loss of the positive and negative features and the distribution difference between the positive features and the routine, so that the features can be distinguished while enriching the feature information;
3. the invention provides more accurate and personalized recommendation suggestions for the user by being superior to the existing methods on a plurality of real data.
Drawings
In order to more clearly illustrate the embodiments of the present application or technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present invention, and other drawings can be obtained by those skilled in the art according to the drawings.
Fig. 1 is a schematic flow chart of a cold-start recommendation oriented anti-self-coding migration learning method of the present invention.
Detailed Description
In order to make the technical solutions of the present invention better understood, those skilled in the art will now describe the present invention in further detail with reference to the accompanying drawings.
The invention provides a cold-start recommendation oriented anti-self-coding transfer learning method as shown in figure 1, which comprises the following steps:
s1, constructing a hypergraph, modeling interaction information of a user and an article, and designing a positive hypergraph and a negative hypergraph according to an original hypergraph;
s1.1, preprocessing a user scoring matrix, deleting articles with the score of 0, and reserving the rest articles as interactive items;
s1.2, according to the following steps that 9:1 into warm users and cold start users, and constructing a hypergraph according to the interactive information of the warm users and the articles, specifically
Figure BDA0003798787200000071
And the form of the hypergraph is as follows:
Figure BDA0003798787200000072
wherein n is w Representing the number of warm users, m represents the total number of items, the value of R (i, j) is 1 representing that the warm user i has interaction with the item j, otherwise the value of R (i, j) is 0;
s1.3, positive superscript graph R + Equal to the original hypergraph R, the positive hypergraph represents the preference items of the user, and the items with the value of R (i, j) of 0 are randomly selected from the original hypergraph R to construct a negative hypergraph R - ；
The form of the negative hypergraph is as follows:
Figure BDA0003798787200000081
s2, designing a multi-layer perceptron network to obtain conventional feature representations of warm users and cold start users;
s2.1, constructing a multilayer perceptron neural network to obtain the conventional characteristics of the warm user and the cold start user, wherein the specific formula is as follows:
Figure BDA0003798787200000082
where h (-) represents a fully connected neural network,
Figure BDA0003798787200000083
representing the characteristics of users, the trust relations of the users exist in the original data, the trust relations are used as the trust relations, n represents the total number of the users, d represents the dimension of the characteristics, phi represents the trainable network weight and has
U＝[U w ,U c ]
Figure BDA0003798787200000084
U w And U c Representing the original characteristics of a warm user and a cold start user respectively,
Figure BDA0003798787200000085
and
Figure BDA0003798787200000086
conventional characterization representative of warm and cold start users, respectively;
and S3, constructing a hypergraph self-encoder for the positive hypergraph and the negative hypergraph respectively, acquiring positive and negative feature representations of the warm user and the article, and regarding the positive features as source data and the conventional features as target data. In addition, positive and negative hypergraphs are reconstructed and used for storing the associated information between the user and the article;
s3.1, the relation between the user and the article is represented by a hypergraph, and the hypergraph self-encoder is used for learning high-level feature representation of the warm user and the article on the hypergraph, wherein the specific formula is as follows:
Figure BDA0003798787200000087
wherein f is + Is a hypergraph self-encoder, T is a characteristic of the article,
Figure BDA0003798787200000088
in order to reconstruct the hypergraph,
Figure BDA0003798787200000089
is a positive feature of a warm user,
Figure BDA00037987872000000810
is a positive feature of the article;
s3.2, reconstructing a positive hypergraph by using the positive characteristics of the warm user and the characteristics of the article acquired in the step S3.1, wherein the loss of the reconstructed positive hypergraph is represented as:
Figure BDA0003798787200000091
wherein
Figure BDA0003798787200000092
Is a binary cross entropy function.
S3.3, learning high-level feature representation of warm users and articles by using a hypergraph self-encoder on the negative hypergraph, wherein a specific formula is as follows:
Figure BDA0003798787200000093
wherein f is - Is a hypergraph autoencoder.
Figure BDA0003798787200000094
In order to reconstruct the hypergraph,
Figure BDA0003798787200000095
is a negative feature of a warm user,
Figure BDA0003798787200000096
is a negative characteristic of the article;
s3.4, reconstructing a negative hypergraph by using the negative characteristics of the warm user and the characteristics of the article acquired in the step S3.3, wherein the loss of the reconstructed negative hypergraph can be expressed as:
Figure BDA0003798787200000097
s4, constructing a matching discriminator, distributing pseudo labels for the positive and negative characteristics of the warm user, and minimizing the classification loss of the positive and negative characteristics of the warm user and the distribution difference between the positive characteristics and the conventional characteristics;
s4.1 Positive characteristics for warming users respectively
Figure BDA0003798787200000098
And negative characteristics
Figure BDA0003798787200000099
Assigning pseudo-labels, i.e. setting
Figure BDA00037987872000000910
Assign a pseudo label as
Figure BDA00037987872000000911
Wherein n is w The number of warm users, d is the characteristic dimension;
the positive characteristics comprise preference information of the user, and the negative characteristics comprise information which is not interested by the user;
s4.2, positive feature to bridge Warm user and regular feature, i.e. settings
Figure BDA00037987872000000912
Assign a pseudo label as
Figure BDA00037987872000000913
S4.3, designing a matching discriminator, and minimizing the classification loss of the positive and negative characteristics of the warm user and the distribution difference between the positive characteristics and the conventional characteristics according to the following formula:
Figure BDA0003798787200000101
where a is a trainable parameter which is,
Figure BDA0003798787200000102
and
Figure BDA0003798787200000103
respectively represent
Figure BDA0003798787200000104
And
Figure BDA0003798787200000105
the (j) th row vector of (a),
Figure BDA0003798787200000106
and
Figure BDA0003798787200000107
respectively represent
Figure BDA0003798787200000108
And
Figure BDA0003798787200000109
value of j row of (1), G y (. Is) a classifier, G c (ii) a characteristic matcher,
Figure BDA00037987872000001010
is a loss function of the classifier and is,
Figure BDA00037987872000001011
representing the penalty function between domains;
s4.4, by step S4.2, the positive and negative characteristics of the warm user are separated, the positive characteristics are connected with the regular characteristics, thereby bridging the regular characteristics of the cold-start user and the item characteristic image recommendation, introducing a gradient back propagation layer, following the following formula:
Figure BDA00037987872000001012
s4.5, integrating the steps, designing an overall loss function of the model for training, wherein the formula is as follows:
Figure BDA00037987872000001013
where β and η are adjustable parameters to balance the weights between the three loss functions;
s4.6, repeating the steps S4.3 and S4.4 until convergence, and connecting the positive characteristic and the conventional characteristic of the warm user;
the positive characteristics of the warm user and the positive characteristics of the article are connected through a positive hypergraph, the conventional characteristics of the warm user and the conventional characteristics of the cold start user have similar distribution, and the relationship between the conventional characteristics of the cold start user and the positive characteristics of the article is observed and recorded;
s5, calculating Euclidean distances between the conventional features and the article features of the cold-start user, sequencing, recommending Top-K articles to the user, and calculating precision, recall rate, NDCG and hit rate;
s5.1, discovering the relation between the cold start user and the articles, calculating the Euclidean distance between the conventional characteristics and the article characteristics of the cold start user, recommending Top-K articles to the user, and evaluating the performance of the method by adopting the following four indexes:
A. precision: the ratio of the total number of the items in the recommendation list is calculated, all users are averaged, and the larger the value is, the better the value is, the calculation formula is as follows:
Figure BDA0003798787200000111
wherein N is ts Is coldNumber of active users, L i Is the ith cold start user's true favorite item,
Figure BDA0003798787200000112
is the recommended Top-K items;
B. and (4) recall rate: as the proportion of correctly recommended articles to the total number of articles to be recommended, the larger the value is, the better the value is, and the calculation formula is as follows:
Figure BDA0003798787200000113
c, NDCG: for measuring the superiority of the recommendation list, when the result with high relevance appears at a more advanced position, the higher the index is, the calculation formula is as follows:
Figure BDA0003798787200000114
wherein r is i Is the correlation of the ith item, the user prefers the item i then r i A value of 1, otherwise 0, while the value of IDCG ensures that the value of NDCG can be between 0 and 1;
D. the hit rate: as a commonly used index for measuring the recall ratio, the larger the value is, the better the value is, and the calculation formula is as follows:
Figure BDA0003798787200000115
s5.2, using the Ciao data set for a user cold start recommendation problem, collecting scores and opinions of the user on various products from the Ciao official website, deleting the items with the score of 0, and finally selecting 2153 users and 8000 products, wherein the user is subjected to the following steps of 9:1 divided into warm users and cold start users;
s5.3, comparing the Top-20 prediction result on the Ciao with CMF (David cortex.2018. Cold-start criteria in collectible matrix factorization. ArXiv prediction arXiv:1809.00366 (2018)) and Heater (Ziwei Zhu, shahin Sefati, parsa Saadatpana, and James conduit.2020. Recommendation for new users and new items via random tracking and mixing-of-experiments transformation. In ACM SIGIR Conference Research and Development Information retrieval.1121-1130);
TABLE 1
Figure BDA0003798787200000121
The experimental results are shown in table 1, and the algorithm 1 corresponds to the verification results of the algorithm provided by the invention; the algorithm 2 corresponds to the verification result of the CMF; the algorithm 3 corresponds to the verification result of Heater; as can be seen from Table 1, the present invention is better than other algorithms in most evaluation indexes;
while certain exemplary embodiments of the present invention have been described above by way of illustration only, it will be apparent to those of ordinary skill in the art that the described embodiments may be modified in various different ways without departing from the spirit and scope of the present invention. Accordingly, the drawings and description are illustrative in nature and are not to be construed as limiting the scope of the invention.

Claims (6)

1. A cold-start recommendation-oriented antagonistic self-coding transfer learning method is characterized by comprising the following steps: the method comprises the following steps:
s1, constructing a hypergraph, modeling interactive information of a user and an article, and designing a positive hypergraph and a negative hypergraph according to the original hypergraph;
s2, designing a multi-layer perceptron network to obtain conventional feature representations of warm users and cold start users;
and S3, constructing a hypergraph self-encoder for respectively using the positive hypergraph and the negative hypergraph, acquiring positive and negative feature representations of the warm user and the warm object, and regarding the positive features as source data and the conventional features as target data. In addition, positive and negative hypergraphs are reconstructed and used for storing the associated information between the user and the article;
s4, constructing a matching discriminator, distributing pseudo labels for the positive and negative characteristics of the warm user, and minimizing the classification loss of the positive and negative characteristics of the warm user and the distribution difference between the positive characteristics and the conventional characteristics;
s5, calculating Euclidean distances between the conventional features and the article features of the cold-start user, sequencing, recommending Top-K articles to the user, and calculating precision, recall rate, NDCG and hit rate.
2. The method for learning against self-coding migration facing to cold start recommendation according to claim 1, wherein: the step S1 includes:
s1.1, preprocessing a user scoring matrix, deleting articles with the score of 0, and reserving the rest articles as interactive items;
s1.2, enabling a user to perform the following steps according to the following steps of 9:1 into warm users and cold start users, and constructing a hypergraph according to the interactive information of the warm users and the articles, specifically
Figure FDA0003798787190000011
And the form of the hypergraph is as follows:
Figure FDA0003798787190000012
wherein n is w Representing the number of warm users, m represents the total number of items, the value of R (i, j) is 1 representing that the warm user i has interaction with the item j, otherwise the value of R (i, j) is 0;
s1.3, positive superscript R + Equal to the original hypergraph R, the positive hypergraph represents the preference items of the user, and the items with the value of R (i, j) of 0 are randomly selected from the original hypergraph R to construct a negative hypergraph R - ；
The form of the negative hypergraph is then as follows:
Figure FDA0003798787190000021
。
3. the method for learning against self-coding migration for cold start recommendation according to claim 1, wherein: the step S2 includes:
s2.1, constructing a multilayer perceptron neural network to obtain the conventional characteristics of the warm user and the cold start user, wherein the specific formula is as follows:
Figure FDA0003798787190000022
where h (-) denotes a fully connected neural network,
Figure FDA0003798787190000023
representing the characteristics of users, the trust relations of the users exist in the original data, the trust relations are used as the trust relations, n represents the total number of the users, d represents the dimension of the characteristics, phi represents the trainable network weight and has
U＝[U w ,U c ]
Figure FDA0003798787190000024
U w And U c Representing the original characteristics of a warm user and a cold start user respectively,
Figure FDA0003798787190000025
and
Figure FDA0003798787190000026
conventional characterizations representing warm and cold start users, respectively.
4. The method for learning against self-coding migration for cold start recommendation according to claim 1, wherein: the step S3 includes:
s3.1, the relation between the user and the article is represented by a hypergraph, and the hypergraph self-encoder is used for learning high-level feature representation of the warm user and the article on the hypergraph, wherein the specific formula is as follows:
Figure FDA0003798787190000027
wherein f is + Is a hypergraph self-encoder, T is the characteristic of the article,
Figure FDA0003798787190000028
in order to reconstruct the hypergraph,
Figure FDA0003798787190000029
is a positive feature of a warm user,
Figure FDA00037987871900000210
is a positive feature of the article;
s3.2, reconstructing a positive hypergraph by using the positive characteristics of the warm user and the characteristics of the article acquired in the step S3.1, wherein the loss of the reconstructed positive hypergraph is represented as:
Figure FDA0003798787190000031
wherein
Figure FDA0003798787190000032
Is a binary cross entropy function.
S3.3, learning high-level characteristic representation of warm users and articles by using a hypergraph self-encoder on the negative hypergraph, wherein a specific formula is as follows:
Figure FDA0003798787190000033
wherein f is - Is a hypergraph self-encoder.
Figure FDA0003798787190000034
In order to reconstruct the hypergraph,
Figure FDA0003798787190000035
is a negative feature of a warm user,
Figure FDA0003798787190000036
is a negative characteristic of the article;
s3.4, reconstructing a negative hypergraph by using the negative characteristics of the warm user and the characteristics of the article acquired in the step S3.3, wherein the loss of the reconstructed negative hypergraph can be expressed as:
Figure FDA0003798787190000037
。
5. the method for learning against self-coding migration for cold start recommendation according to claim 1, wherein: the step S4 includes:
s4.1 Positive characteristics for warming users respectively
Figure FDA0003798787190000038
And negative characteristics
Figure FDA0003798787190000039
Assigning pseudo-labels, i.e. setting
Figure FDA00037987871900000310
Assign a pseudo label as
Figure FDA00037987871900000311
Wherein n is w The number of warm users, d is the characteristic dimension;
the positive characteristics contain preference information of the user, and the negative characteristics contain information which is not interesting to the user;
s4.2, positive feature to bridge Warm user and regular feature, i.e. settings
Figure FDA00037987871900000312
Assign a pseudo label as
Figure FDA00037987871900000313
S4.3, designing a matching discriminator, and minimizing the classification loss of the positive and negative characteristics of the warm user and the distribution difference between the positive characteristics and the conventional characteristics according to the following formula:
Figure FDA0003798787190000041
where a is a trainable parameter which is,
Figure FDA0003798787190000042
and
Figure FDA0003798787190000043
respectively represent
Figure FDA0003798787190000044
And
Figure FDA0003798787190000045
the (j) th row vector of (a),
Figure FDA0003798787190000046
and
Figure FDA0003798787190000047
respectively represent
Figure FDA0003798787190000048
And
Figure FDA0003798787190000049
value of j row of (1), G y (. Is a classifier, G) c (v) a characteristic matcher for matching the characteristics of the target,
Figure FDA00037987871900000410
is a loss function of the classifier and,
Figure FDA00037987871900000411
representing the penalty function between domains;
s4.4, by step S4.2, the positive and negative characteristics of the warm user are separated, the positive characteristics are connected with the regular characteristics, thereby bridging the regular characteristics of the cold start user and the item characteristic image recommendation, introducing a gradient back propagation layer, following the following formula:
Figure FDA00037987871900000412
s4.5, integrating the steps, designing a total loss function of the model for training, wherein the formula is as follows:
Figure FDA00037987871900000413
where β and η are adjustable parameters to balance the weights between the three loss functions;
s4.6, repeating the steps S4.3 and S4.4 until convergence, and connecting the positive characteristic and the conventional characteristic of the warm user;
the positive characteristics of the warm user and the positive characteristics of the article are connected through a positive supergraph, the normal characteristics of the warm user and the normal characteristics of the cold start user have similar distribution, and the normal characteristics of the cold start user and the positive characteristics of the article are observed and recorded.
6. The method for learning against self-coding migration for cold start recommendation according to claim 1, wherein: the step S5 includes:
s5.1, discovering the relation between the cold start user and the articles, calculating the Euclidean distance between the conventional characteristics and the article characteristics of the cold start user, recommending Top-K articles to the user, and evaluating the performance of the method by adopting the following four indexes:
A. precision: the ratio of the total number of the items in the recommendation list is calculated, all users are averaged, and the larger the value is, the better the value is, the calculation formula is as follows:
Figure FDA0003798787190000051
wherein N is ts Is the number of cold start users, L i Is the ith cold start user's true favorite item,
Figure FDA0003798787190000052
is the recommended Top-K items;
B. the recall ratio is as follows: as the proportion of correctly recommended articles to the total number of articles to be recommended, the larger the value is, the better the value is, and the calculation formula is as follows:
Figure FDA0003798787190000053
c, NDCG: for measuring the superiority of the recommendation list, when the result with high relevance appears at a more front position, the higher the index is, the calculation formula is as follows:
Figure FDA0003798787190000054
wherein r is i Is the correlation of the ith item, the user prefers the item i then r i The value of the sum of the values is 1,otherwise 0, and the value of IDCG ensures that the value of NDCG can be between 0 and 1;
D. the hit rate is as follows: as a commonly used index for measuring the recall ratio, the larger the value is, the better the value is, and the calculation formula is as follows:
Figure FDA0003798787190000055
。
CN202210976839.3A 2022-08-15 2022-08-15 Cold start recommendation-oriented countermeasure self-coding migration learning method Active CN115422442B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210976839.3A CN115422442B (en) 2022-08-15 2022-08-15 Cold start recommendation-oriented countermeasure self-coding migration learning method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210976839.3A CN115422442B (en) 2022-08-15 2022-08-15 Cold start recommendation-oriented countermeasure self-coding migration learning method

Publications (2)

Publication Number Publication Date
CN115422442A true CN115422442A (en) 2022-12-02
CN115422442B CN115422442B (en) 2024-01-19

Family

ID=84198973

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210976839.3A Active CN115422442B (en) 2022-08-15 2022-08-15 Cold start recommendation-oriented countermeasure self-coding migration learning method

Country Status (1)

Country Link
CN (1) CN115422442B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106600065A (en) * 2016-12-16 2017-04-26 中山大学 Individualized learning path extraction and jointing method and system based on directed hypergraph
US20180276540A1 (en) * 2017-03-22 2018-09-27 NextEv USA, Inc. Modeling of the latent embedding of music using deep neural network
US20190325343A1 (en) * 2018-04-19 2019-10-24 National University Of Singapore Machine learning using partial order hypergraphs
CN111881350A (en) * 2020-07-23 2020-11-03 清华大学 Recommendation method and system based on mixed graph structured modeling
CN113672693A (en) * 2021-08-23 2021-11-19 东北林业大学 Label recommendation method of online question and answer platform based on knowledge graph and label association

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106600065A (en) * 2016-12-16 2017-04-26 中山大学 Individualized learning path extraction and jointing method and system based on directed hypergraph
US20180276540A1 (en) * 2017-03-22 2018-09-27 NextEv USA, Inc. Modeling of the latent embedding of music using deep neural network
US20190325343A1 (en) * 2018-04-19 2019-10-24 National University Of Singapore Machine learning using partial order hypergraphs
CN111881350A (en) * 2020-07-23 2020-11-03 清华大学 Recommendation method and system based on mixed graph structured modeling
CN113672693A (en) * 2021-08-23 2021-11-19 东北林业大学 Label recommendation method of online question and answer platform based on knowledge graph and label association

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
毛明松;张富国;: "基于多重图排序的用户冷启动推荐方法", 计算机工程, no. 05 *

Also Published As

Publication number Publication date
CN115422442B (en) 2024-01-19

Similar Documents

Publication Publication Date Title
Yu et al. A cross-domain collaborative filtering algorithm with expanding user and item features via the latent factor space of auxiliary domains
Zhou et al. Userrec: A user recommendation framework in social tagging systems
CN108460619B (en) Method for providing collaborative recommendation model fusing explicit and implicit feedback
CN108647996A (en) A kind of personalized recommendation method and system based on Spark
Liu et al. Exploiting aesthetic preference in deep cross networks for cross-domain recommendation
CN109241366B (en) Hybrid recommendation system and method based on multitask deep learning
CN108345697A (en) Wisdom course towards group of college students recommends method, system and storage medium
Xie et al. A hybrid semantic item model for recipe search by example
CN113918833A (en) Product recommendation method realized through graph convolution collaborative filtering of social network relationship
Ng et al. CrsRecs: a personalized course recommendation system for college students
Liang et al. Personalized recommender system based on item taxonomy and folksonomy
CN112818256B (en) Recommendation method based on neural collaborative filtering
Wu et al. Modeling uncertainty driven curiosity for social recommendation
CN102866997B (en) The treating method and apparatus of user data
CN115422442A (en) Cold-start recommendation-oriented anti-self-coding migration learning method
Nazari et al. Scalable and data-independent multi-agent recommender system using social networks analysis
CN113377973B (en) Article recommendation method based on countermeasures hash
Banouar et al. Enriching SPARQL queries by user preferences for results adaptation
Kirsch Social information retrieval
Liu et al. Deep cross networks with aesthetic preference for cross-domain recommendation
Melamed et al. MarCol: A market-based recommender system
Hu et al. Utilizing users' tipping points in E-commerce Recommender systems
Liu et al. Multi-domain collaborative recommendation with feature selection
Mohan et al. Recommendation system in business intelligence solutions for grocery shops: Challenges and perspective
Chao et al. Collaborative Filtering and Leaders' Advice Based Recommendation System for Cold Start Users

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