CN116051224B - Cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction - Google Patents
Cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction Download PDFInfo
- Publication number
- CN116051224B CN116051224B CN202211313565.6A CN202211313565A CN116051224B CN 116051224 B CN116051224 B CN 116051224B CN 202211313565 A CN202211313565 A CN 202211313565A CN 116051224 B CN116051224 B CN 116051224B
- Authority
- CN
- China
- Prior art keywords
- user
- target domain
- domain
- representation
- cold start
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 51
- 230000000295 complement effect Effects 0.000 title claims abstract description 35
- 238000000605 extraction Methods 0.000 title claims abstract description 32
- 238000013508 migration Methods 0.000 title claims abstract description 23
- 230000005012 migration Effects 0.000 title claims abstract description 23
- 238000013507 mapping Methods 0.000 claims description 78
- 230000004927 fusion Effects 0.000 claims description 54
- 239000013598 vector Substances 0.000 claims description 53
- 230000003993 interaction Effects 0.000 claims description 23
- 238000005457 optimization Methods 0.000 claims description 9
- 230000002452 interceptive effect Effects 0.000 claims description 8
- 230000007246 mechanism Effects 0.000 claims description 7
- 238000005070 sampling Methods 0.000 claims description 7
- 238000012549 training Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 description 25
- 238000002474 experimental method Methods 0.000 description 16
- 230000006872 improvement Effects 0.000 description 10
- 230000000694 effects Effects 0.000 description 9
- 230000008569 process Effects 0.000 description 8
- 238000005065 mining Methods 0.000 description 7
- 238000004422 calculation algorithm Methods 0.000 description 6
- 238000011156 evaluation Methods 0.000 description 6
- 230000007547 defect Effects 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 238000002679 ablation Methods 0.000 description 4
- 238000000354 decomposition reaction Methods 0.000 description 3
- 238000012546 transfer Methods 0.000 description 3
- 230000006399 behavior Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 101100136092 Drosophila melanogaster peng gene Proteins 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000001939 inductive effect Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 125000000391 vinyl group Chemical group [H]C([*])=C([H])[H] 0.000 description 1
- 229920002554 vinyl polymer Polymers 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Physics & Mathematics (AREA)
- Economics (AREA)
- General Business, Economics & Management (AREA)
- Strategic Management (AREA)
- Marketing (AREA)
- Development Economics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention relates to the technical field of cross-domain recommendation, in particular to a cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction, which comprises the following steps: acquiring a plurality of commodities to be recommended of a cold start user existing in a target domain in the target domain; inputting the source domain user representation of the cold start user into the trained cross-domain recommendation model, and outputting a corresponding target domain end user representation; calculating a target domain prediction score of the corresponding cold start user for the commodity to be recommended based on the target domain end user representation of the cold start user in combination with the corresponding target domain commodity representation to be recommended; and taking one or more commodities to be recommended with highest target domain prediction scores as recommendation results of the cold start user. The cross-domain recommendation method can realize complementation of the user commonality attribute and the personality attribute, and can fully mine the knowledge of the target domain.
Description
Technical Field
The invention relates to the technical field of cross-domain recommendation, in particular to a cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction.
Background
The recommendation system is widely applied to various business occasions, such as an electronic commerce system, a social network system and the like, and plays an increasingly important role in the background that the information overload problem is more and more prominent because the recommendation system can screen out a proper part from massive information to provide recommendation services for users. In recent years, research work of recommended systems has attracted a large number of researchers from industry and academia. However, most recommendation system models do not solve the data sparsity problem well, so that the system cannot provide satisfactory recommendation performance for cold start users (i.e. new users) all the time. As an effective data sparsity problem solution, the cross-domain recommendation technique alleviates the data sparsity problem of the target domain by migrating source domain knowledge, which is relatively rich in user information, into the target domain.
In order to better migrate knowledge of a source domain to a target domain, it is proposed in the prior art to learn a common mapping function between the source domain and the target domain for all users based on the assumption of "relationship sharing between preferences of all users in the source domain and the target domain" to achieve migration of knowledge in the source domain to the target domain. However, this assumption is not reasonable, because the relationship between the preferences of different users in the source domain and the target domain is not completely consistent, and this method of learning a common mapping function between the source domain and the target domain to perform knowledge migration is difficult to accurately reflect the preference information of each user, so the above-mentioned existing cross-domain recommendation system cannot reflect the personalized preferences of the users. For this problem, the existing PTUPCDR model proposes to provide a personalized mapping function for each user, and considers different preference information of different users, so that the defect of a method based on a public mapping function is overcome better, and better cross-domain recommendation performance is realized.
However, the method based on the public mapping function only focuses on the public attribute among users, and lacks of explicit modeling on the individual attribute of different users. While the PTUPCDR model considers different preference information of different users, the method based on the personalized mapping function only focuses on the personalized attribute of single users and lacks explicit modeling of the attribute common to all users. Therefore, neither the method based on the common mapping function nor the method based on the personalized mapping function is misdirected to modeling of the user, and modeling alone cannot optimize the model performance in any aspect, which results in poor comprehensiveness of the cross-domain recommendation. Meanwhile, the existing cross-domain recommendation algorithm focuses on how to better migrate knowledge of a source domain to a target domain too much, but neglects exploration of the knowledge of the target domain, ignores the effect of certain information of the target domain on solving the cold start problem, and if interaction data of certain users in the source domain are insufficient, the learned knowledge of the source domain is relatively insufficient, so that the knowledge of source domain migration is naturally limited in helping the cold start problem of the target domain, and further the accuracy of cross-domain recommendation is poor. Therefore, how to design a method capable of improving the comprehensiveness and accuracy of cross-domain commodity recommendation is a technical problem to be solved.
Disclosure of Invention
Aiming at the defects in the prior art, the invention aims to solve the technical problems that: how to provide a cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction, so as to realize complementation of user commonality attribute and personality attribute, and fully mine knowledge of the target domain, thereby improving comprehensiveness and accuracy of cross-domain commodity recommendation.
In order to solve the technical problems, the invention adopts the following technical scheme:
a cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction comprises the following steps:
s1: acquiring a plurality of commodities to be recommended of a cold start user existing in a target domain in the target domain;
s2: inputting the source domain user representation of the cold start user into the trained cross-domain recommendation model, and outputting a corresponding target domain end user representation; the cross-domain recommendation model comprises a complementary mapping module and a feature extraction module;
the complementary mapping module is used for generating a corresponding target domain personality representation and a target domain commonality representation based on the source domain user representation of the cold start user; then fusing the corresponding target domain personality representation and target domain commonality representation to obtain a target domain fusion user representation of the cold start user;
The feature extraction module is used for sampling a plurality of user sets from the target domain user sets to serve as key sets; then combining the target domain fusion user representations of the cold start users, and calculating the attention scores of the target domain fusion user representations corresponding to the cold start users for the users in the corresponding key set by using an attention mechanism; calculating target domain fusion user expression of the cold start user by combining the attention scores of the users in the corresponding key sets, and representing target domain feature vectors of the corresponding key sets; finally, fusing target domain fusion user expression of the cold start user with target domain feature vectors of each key set to obtain target domain enhancement user expression of the cold start user;
fusing target domain fusion user representation and target domain enhancement user representation of a cold start user by a cross-domain recommendation model to obtain corresponding target domain end user representation;
s3: calculating a target domain prediction score of the corresponding cold start user for the commodity to be recommended based on the target domain end user representation of the cold start user in combination with the corresponding target domain commodity representation to be recommended;
s4: and taking one or more commodities to be recommended with highest target domain prediction scores as recommendation results of the cold start user.
Preferably, in step S2, the target domain personality representation of the cold start user is generated by:
s201: commodity interaction sequence for cold start user in source domainInputting a source domain recommendation model, and outputting an embedded vector representation of the obtained commodity interaction sequence>n represents the number of interactive items;
wherein:indicating that the user is at t k The embedded vectors of the goods are interacted at each moment; />Indicating that the user is at t k Goods interacted at each moment; />Representing a source domain recommendation model;
s202: computing cold start user-to-commodity interaction sequence based on attention mechanismThe attention score of each commodity in the commodity interaction sequence is further represented by an embedded vector of the commodity interaction sequence>The embedded vectors of the commodities are combined with the corresponding attention scores to generate user characteristics based on the commodities;
wherein:representing commodity-based user characteristics; alpha j Representing a cold start user u i Interactive sequence for commodityCommodity v of China j Is the concentration score of (1), wherein +.>Attn (·) represents the attention network, +.>Parameters representing an attention network; />Embedded vector representation representing a merchandise interaction sequence +.>An embedded vector of the commodity; a, a l Representing the attention score prior to normalization, i.e. the output of the attention network; />Representing a commodity;
S203: taking the commodity-based user characteristics as the input of the meta-network, and taking the output of the meta-network as the weight parameter of the personalized characteristic mapping model; then the source domain user representation of the cold start user is used as the input of the personality characteristic mapping model, and the output of the personality characteristic mapping model is used as the target domain personality representation of the cold start user;
wherein:the output of the meta-network, namely the weight parameter of the personality characteristic mapping model; />Representing commodity-based user characteristics; f (f) meta (-) represents a meta-network; epsilon represents a meta-network parameter; />The output of the personality characteristic mapping model is represented, namely the personality of the target domain of the cold start user is represented; f (f) p Representing a personality trait map model; />Representing a cold start user u i Is a source domain user representation of (1).
Preferably, in step S2, the source domain user representation of the cold start user is used as an input of the commonality feature mapping model, and the output of the commonality feature mapping model is used as the target domain commonality representation of the cold start user;
wherein:the output of the common feature mapping model is represented, namely the target domain common representation of the cold start user; />Representing a cold start user u i Source domain user representation of (2); f (f) com (. Cndot.) represents a commonality feature mapping model; ω represents model parameters.
Preferably, in step S2, the target domain fusion user representation of the cold start user is calculated by the following formula;
wherein:a target domain fusion user representation representing a cold boot user; />A target domain personality representation representing a cold boot user; />A target domain commonality representation representing a cold-start user.
Preferably, in step S2, the attention score of the target domain fusion user representation of the cold start user for the corresponding user in the corresponding key set is calculated by the following formula;
wherein:representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Middle user->Is a fraction of the attention of (2); />A target domain fusion user representation representing a cold boot user; all represent trainable parameters; />Representing a join operation; />Representing a vector representation corresponding to the user.
Preferably, in step S2, the target domain feature vector of the cold start user, which represents the target domain feature vector for the key set, is calculated by the following formula;
wherein:representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Is defined by the target domain feature vector of (a); />Representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key setMiddle user- >Is a fraction of the attention of (2); />Representing the trainable parameters.
Preferably, in step S2, the target domain enhanced user representation of the cold start user is calculated by the following formula;
wherein:a target domain enhanced user representation representing a cold boot user; />Representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Is defined by the target domain feature vector of (a); />Representing the trainable parameters.
Preferably, in step S2, the target domain end user representation of the cold start user is calculated by the following formula;
wherein:a target domain end user representation representing a cold boot user; />A target domain enhanced user representation representing a cold boot user; />The target domain representing the cold-boot user fuses the user representations.
Preferably, in step S3, the target domain prediction score of the commodity to be recommended is calculated by the following formula;
wherein:target domain end user representation representing cold boot user +.>Commodity representation with target domain to be recommendedIs the inner product of (i) cold start user u i For commodity v to be recommended j Target domain prediction scores of (2).
Preferably, in step S2, when the cross-domain recommendation model is trained, an optimization target is calculated by the following formula;
wherein:expressing an optimization target during cross-domain recommendation model training; / >Representing a set of true scores of public users of the source domain and the target domain in the target domain; />Representing a cold start user u i For commodity v to be recommended j Target domain prediction scores for (2); r is (r) ij Representing a cold start user u i For commodity v to be recommended j Target domain true score of (2).
The cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction has the following beneficial effects:
according to the method and the device for recommending the cross-domain commodity, the complementary mapping module of the cross-domain recommendation model generates the target domain personalized representation and the target domain commonality representation based on the source domain user representation of the cold-start user, and then the target domain fusion user representation is obtained through fusion, so that the personalized information and the commonality information of the user can be considered simultaneously, the complementation of the personalized attribute and the commonality attribute of the user can be realized, the characteristics of the user in the target domain can be more comprehensively described, and the comprehensiveness of the cross-domain commodity recommendation can be improved.
According to the method, a plurality of key sets are sampled from the target domain user set by the cross-domain recommendation model of the cross-domain recommendation model, the attention score of each user in the key set and the target domain feature vector of the key set are calculated by combining the target domain fusion user representation and the attention mechanism, and then the target domain feature vector of each key set is fused to obtain the target domain enhancement user representation, so that knowledge of the target domain can be fully mined by combining a relational network, the defect that the existing cross-domain recommendation model ignores knowledge mining of the target domain is overcome, and further performance of the cross-domain recommendation model can be better improved by extracting proper target domain features, and accuracy of cross-domain commodity recommendation can be improved.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings, in which:
FIG. 1 is a logical block diagram of a cross-domain recommendation method;
FIG. 2 is a network structure diagram of a cross-domain recommendation model;
FIG. 3 is a comparison experiment of JCTCDR model and EMCDR, PTUPCDR with respect to Mean Absolute Error (MAE) performance based on MF model;
FIG. 4 is a graph showing performance comparisons of JCTCDR model and EMCDR, PTUPCDR with respect to Root Mean Square (RMSE) based on MF model;
FIG. 5 is a graph showing Mean Absolute Error (MAE) performance versus JCTCDR model and EMCDR, PTUPCDR based on GMF model;
FIG. 6 is a graph showing performance comparisons of JCTCDR model and EMCDR, PTUPCDR with respect to Root Mean Square (RMSE) based on GMF model;
FIG. 7 is a graph showing the effect of the number of users in a critical set on the Mean Absolute Error (MAE) for various tasks;
FIG. 8 is a graph showing the effect of the number of users in a critical set on Root Mean Square Error (RMSE) for different tasks;
FIG. 9 is a graph showing the effect of the number of critical sets on Mean Absolute Error (MAE) performance for different tasks;
FIG. 10 is a graph of the effect of the number of critical sets on Root Mean Square Error (RMSE) performance for different tasks.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. It will be apparent that the described embodiments are some, but not all, embodiments of the invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the invention, as presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that: like reference numerals and letters denote like items in the following figures, and thus once an item is defined in one figure, no further definition or explanation thereof is necessary in the following figures. In the description of the present invention, it should be noted that, directions or positional relationships indicated by terms such as "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., are directions or positional relationships based on those shown in the drawings, or are directions or positional relationships conventionally put in use of the inventive product, are merely for convenience of describing the present invention and simplifying the description, and are not indicative or implying that the apparatus or element to be referred to must have a specific direction, be constructed and operated in a specific direction, and thus should not be construed as limiting the present invention. Furthermore, the terms "first," "second," "third," and the like are used merely to distinguish between descriptions and should not be construed as indicating or implying relative importance. Furthermore, the terms "horizontal," "vertical," and the like do not denote a requirement that the component be absolutely horizontal or overhang, but rather may be slightly inclined. As "horizontal" merely means that its direction is more horizontal than "vertical", and does not mean that the structure must be perfectly horizontal, but may be slightly inclined. In the description of the present invention, it should also be noted that, unless explicitly specified and limited otherwise, the terms "disposed," "mounted," "connected," and "connected" are to be construed broadly, and may be, for example, fixedly connected, detachably connected, or integrally connected; can be mechanically or electrically connected; can be directly connected or indirectly connected through an intermediate medium, and can be communication between two elements. The specific meaning of the above terms in the present invention will be understood in specific cases by those of ordinary skill in the art.
The following is a further detailed description of the embodiments:
examples:
the embodiment discloses a cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction.
As shown in fig. 1, a cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction includes:
s1: acquiring a plurality of commodities to be recommended of a cold start user existing in a target domain in the target domain;
s2: the source domain user representation of the cold start user is used as input of a trained cross-domain recommendation model (hereinafter also referred to as JCTCDR model), and the target domain end user representation of the cold start user is output through the cross-domain recommendation model;
referring to fig. 2, the cross-domain recommendation model includes a complementary mapping module and a feature extraction module;
the complementary mapping module is used for generating a corresponding target domain personality representation and a target domain commonality representation based on the source domain user representation of the cold start user; then fusing the corresponding target domain personality representation and target domain commonality representation to obtain a target domain fusion user representation of the cold start user;
the feature extraction module is used for sampling a plurality of user sets from the target domain user sets to serve as key sets; then combining the target domain fusion user representations of the cold start users, and calculating the attention scores of the target domain fusion user representations corresponding to the cold start users for the users in the corresponding key set by using an attention mechanism; calculating target domain fusion user expression of the cold start user by combining the attention scores of the users in the corresponding key sets, and representing target domain feature vectors of the corresponding key sets; finally, fusing target domain fusion user expression of the cold start user with target domain feature vectors of each key set to obtain target domain enhancement user expression of the cold start user;
Fusing target domain fusion user representation and target domain enhancement user representation of a cold start user by a cross-domain recommendation model to obtain corresponding target domain end user representation;
s3: calculating a target domain prediction score of the corresponding cold start user for the commodity to be recommended based on the target domain end user representation of the cold start user in combination with the corresponding target domain commodity representation to be recommended;
s4: the item or items to be recommended with the highest target domain predictive score are used as the (cross-domain) recommendation result of the cold start user (in the target domain).
In this embodiment, the commodity can be recommended to the cold start user by predicting the score through the target domain of the commodity to be recommended.
The online shopping platform is taken as an example for explanation: user a, originally the user of the treasured platform, starts to register a new account with the east platform in preparation for using the east platform. At this time, a certain treasured platform is a source domain, a certain east platform is a target domain, and the user A is a cold start user of the certain east platform (target domain).
If the user selects to directly log in a certain east platform (target domain) through an account of a certain treasured platform (source domain), or transfer related information of the certain treasured platform (source domain) to the certain east platform (target domain), the related system takes a source domain user representation of a user A in the certain treasured platform (source domain) as input through a cross-domain recommendation model in the invention, and outputs a target domain end user representation of the user A in the certain east platform (target domain); and then the system calculates target domain end user representations of the user A and target domain commodity representations of commodities such as clothes, shoes and household appliances on a certain east platform (target domain) to obtain target domain prediction scores of the user A on the commodities such as clothes, shoes and household appliances on the certain east platform (target domain), and further can carry out cross-domain commodity recommendation for the user A through the target domain prediction scores.
According to the method and the device for recommending the cross-domain commodity, the complementary mapping module of the cross-domain recommendation model generates the target domain personalized representation and the target domain commonality representation based on the source domain user representation of the cold-start user, and then the target domain fusion user representation is obtained through fusion, so that the personalized information and the commonality information of the user can be considered simultaneously, the complementation of the personalized attribute and the commonality attribute of the user can be realized, the characteristics of the user in the target domain can be more comprehensively described, and the comprehensiveness of the cross-domain commodity recommendation can be improved.
According to the method, a plurality of key sets are sampled from the target domain user set by the cross-domain recommendation model of the cross-domain recommendation model, the attention score of each user in the key set and the target domain feature vector of the key set are calculated by combining the target domain fusion user representation and the attention mechanism, and then the target domain feature vector of each key set is fused to obtain the target domain enhancement user representation, so that knowledge of the target domain can be fully mined by combining a relational network, the defect that the existing cross-domain recommendation model ignores knowledge mining of the target domain is overcome, and further performance of the cross-domain recommendation model can be better improved by extracting proper target domain features, and accuracy of cross-domain commodity recommendation can be improved.
In the cross-domain recommendation model, two domains, a source domain and a target domain, are typically included, each domain having a respective user set u= { U 1 ,u 2 ,...,u |U| Commodity set V= { V 1 ,v 2 ,...v |V| Sum of corresponding scoring matricesWherein u is i E U represents the ith (1.ltoreq.i.ltoreq.n) user in the user set, v j E, V represents the j (j is less than or equal to 1 and less than or equal to m) commodity in the commodity set,representing user u i For commodity v j Is a score of (2). To distinguish between the different representations of the two domains, the present implementation is identified using the superscript d e { s, t }: the user, commodity and scoring matrix of the source domain are denoted as U, respectively s 、V s And R is s The users, goods and scores of the target domain are denoted as U, respectively t 、V t And R is t 。
The method of the present patent application is based on the implementation of a public user of two domains, defined as U o ＝U s ∩U t . Thus, whether it is a personality-commonality complementary mapping function moduleBoth learning and learning of the user-perceived target domain feature extraction module are based on the common user set U o . At the same time as the time of this,i.e. there is no common commodity between two different fields. In addition, for the cross-domain recommendation model, cold-start users represent those users that are present in the source domain but not in the target domain, denoted +.>The goal of cross-domain recommendation is to exploit the score R of the source domain s And score R of target domain t To assist the cold start user U e U in the target domain c Is recommended by the user.
In this embodiment, the single-domain recommendation models of the source domain and the target domain are independently trained by using the scoring matrix information of the source domain and the target domain, respectively. For single-domain recommendation models, the most efficient model is the model of hiding factors (from Covington, paul, jay K. Adams, et al deep Neural Networks for YouTube Recommendations), where deep learning and matrix decomposition based models have received more and more attention and have made significant progress in recent years. Since the focus of this patent application is on discussing the framework of cross-domain recommendations, rather than a specific model of single-domain modeling, this embodiment uses f uniformly θ The underlying single-domain recommendation model is represented, where θ is a parameter of the model.
To distinguish between source and target domains, use is made ofRepresenting a source domain recommendation model, use +.>Representing a target domain recommendation model. In the hidden factor model, each user and commodity hidden space is assigned a distributed representation, also known as an embedded vector. This patent application uses->And->Embedded vector representation of the ith user and jth commodity representing the source domain using +.>And->An embedded vector representation of the ith user and jth commodity representing the target domain, where d represents the vector embedded dimension. The probability of scoring the jth item for the ith user may be modeled as follows:
Wherein: p (r) ij |u i ,v j ,σ 2 ) Probabilistic modeling representing scoresAs the average value, with sigma 2 Is a gaussian distribution of variance; sigma (sigma) 2 Is a super parameter; />Representing an embedded vector u v And v j Inner product of f θ The form of (c) depends on the specific hidden factor model.
Optimization of the model parameter θ can be performed by the following formula:
in the method, in the process of the invention,representing model predictive scores; r is (r) ij Representing the true score.
And training respective single-domain recommendation models in the source domain and the target domain respectively through the formulas.
The method based on the public mapping function only pays attention to public attributes among users, and lacks explicit modeling of the attribute of different user personalities; whereas the personalized mapping function-based method focuses only on the personalized attributes of individual users, lacks explicit modeling of common attributes of all users. Therefore, modeling individual characters or commonalities alone may result in suboptimal results, and combining the two together, and modeling complementary to each other, is more likely to achieve better results. Based on this, the present patent application proposes a personality-commonality complementary mapping module (i.e., complementary mapping module) to complementarily model personality attributes and common attributes of users.
In a specific implementation process, for user personalized mapping representation, the patent application refers to the method of using a PTUPCDR model (from Zhu, yongchun, zhenwei Tang, yuman Liu, et al, personalized Transfer of User Preferences for Cross-domain Recommendation) to obtain user characteristics according to an interaction sequence of a user on commodities, then input the user characteristics into a meta-network to obtain an output weight vector, and then use the weight vector as a parameter of a mapping function to achieve the purpose of personalized mapping representation.
Generating a target domain personality representation of the cold start user by:
s201: commodity interaction sequence for cold start user in source domainInputting a source domain recommendation model, and outputting an embedded vector representation of the obtained commodity interaction sequence>n represents the number of interactive items;
wherein:indicating that the user is at t k The embedded vectors of the goods are interacted at each moment; />Indicating that the user is at t k Goods interacted at each moment; />Representing a source domain recommendation model;
s202: computing cold start user-to-commodity interaction sequence based on attention mechanismThe attention score of each commodity in the commodity interaction sequence is further represented by an embedded vector of the commodity interaction sequence>The embedded vectors of the commodities are combined with the corresponding attention scores to generate user characteristics based on the commodities;
wherein:representing commodity-based user characteristics; alpha j Representing a cold start user u i Interactive sequence for commodityCommodity v of China j Is the concentration score of (1), wherein +.>Attn (·) represents the attention network, +.>Parameters representing an attention network; />Embedded vector representation representing a merchandise interaction sequence +.>An embedded vector of the commodity; a, a l Representing the attention score prior to normalization, i.e. the output of the attention network; />Representing a commodity;
the above formula can be intuitively understood that different interactive commodities have different contributions to describing the mobilizable user features useful for the target domain, and the specific contribution is determined by the attention score alpha j To characterize, the greater the score, the greater the contribution.
S203: taking the commodity-based user characteristics as the input of the meta-network, and taking the output of the meta-network as the weight parameter of the personalized characteristic mapping model; then the source domain user representation of the cold start user is used as the input of the personality characteristic mapping model, and the output of the personality characteristic mapping model is used as the target domain personality representation of the cold start user;
wherein:representing the output of a meta-network, i.e. the weights of a personality-characteristic mapping modelA heavy parameter; />Representing commodity-based user characteristics; f (f) meta (. Cndot.) represents a meta-network, which adopts a two-layer perceptron structure; epsilon represents a meta-network parameter; />The output of the personality characteristic mapping model is represented, namely the personality of the target domain of the cold start user is represented; f (f) p Representing a personality trait map model; />Representing a cold start user u i Is a source domain user representation of (1).
In the specific implementation process, although mapped by the user personalized mapping function, the source domain user u i Representation in the target domainIt can reflect user u i But because one of its users shares a pattern of a mapping function alone, resulting in a representation after mapping +.>The personalized features of the user themselves may be too focused, and common features of all users in the source domain that exist in the mapping to the target user are ignored, which may result in loss of part of the common feature information, and thus, the representation of the user mapped to the target domain is not accurate enough.
Based on this problem, the present patent application further proposes a user common feature mapping model for user common feature mining. The source domain user representation of the cold start user is used as input of a commonality feature mapping model, and the output of the commonality feature mapping model is used as target domain commonality representation of the cold start user;
wherein:the output of the common feature mapping model is represented, namely the target domain common representation of the cold start user; />Representing a cold start user u i Source domain user representation of (2); f (f) com (. Cndot.) represents a commonality feature mapping model, and adopts a two-layer perceptron structure; ω represents model parameters.
Calculating a target domain fusion user representation of the cold start user through the following formula;
wherein:a target domain fusion user representation representing a cold boot user; />A target domain personality representation representing a cold boot user; />A target domain commonality representation representing a cold-start user.
According to the method and the device for recommending the cross-domain commodity, the complementary mapping module of the cross-domain recommendation model generates the target domain personalized representation and the target domain commonality representation based on the source domain user representation of the cold-start user, and then the target domain fusion user representation is obtained through fusion, so that the personalized information and the commonality information of the user can be considered simultaneously, the complementation of the personalized attribute and the commonality attribute of the user can be realized, the characteristics of the user in the target domain can be more comprehensively described, and the comprehensiveness of the cross-domain commodity recommendation can be improved.
In a specific implementation process, most of the existing cross-domain recommendation algorithms only focus on how to better and accurately migrate knowledge of a source domain to a target domain, but neglect exploration of knowledge contained in the target domain, and especially when knowledge in the source domain is relatively lacking, the method of migrating knowledge of the source domain cannot bring ideal recommendation performance. Accordingly, the present patent application further proposes a relational network module (i.e. a feature extraction module) for target domain knowledge mining.
Due to source domain usersMapping is represented +.>All information of (2) comes from the migration of source domain knowledge, and does not explicitly consider the characteristics of target domain users, when the source domain knowledge is relatively insufficient per se, the information is directly used + ->Recommending items in the target domain may result in suboptimal results. Thus, extraction of user features in the target domain and integration of the extracted features into the source domain user +.>Representation mapped in the target domain->The middle is necessary.
Inspired by the generalized relational model (Inductive Relation Model), the patent application discloses the extraction of the characteristic information of the target domain userAs a query, the embedded vectors of all users in the target domain are obtained relative to the query vector through the attention mechanism >Weight of +.>Wherein i represents user +.>j represents user +.>Intuitively, weight +.>It can be understood that for obtaining source domain user +.>Embedded vector representation in target domain, target domain user +.>The size of the contribution made is proportional to the weight size. However, in practice, since the number of users in the target domain tends to be large, calculating a weight for each user in the target domain results in a huge amount of calculation, which is impractical for practical use. Therefore, the present patent application adopts a sampling strategy to sample a part of users from the target domain user set as the simulation of the whole target domain user set, and calls the user subset as a 'key set', and then extracts the target domain user characteristics based on the 'key set'. Considering that sampling only a single key set may cause sampling deviation and cannot learn the characteristics of the target domain user more accurately, the present patent application adopts a sampling strategy of multiple key sets, namely, samples multiple different key sets from the target user set to simulate the target domain user set.
Calculating the attention score of the target domain fusion user representation of the cold start user for the corresponding user in the corresponding key set through the following formula;
Wherein:representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Middle user->Is a fraction of the attention of (2); />A target domain fusion user representation representing a cold boot user; /> All represent trainable parameters; />Representing a join operation; />Representing a vector representation corresponding to the user.
Calculating a target domain feature vector of a target domain fusion user representation of the cold start user for the key set by the following formula;
wherein:representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Is defined by the target domain feature vector of (a); />Representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key setMiddle user->Is a fraction of the attention of (2); />Representing the trainable parameters.
Calculating a target domain enhanced user representation of the cold start user by the formula;
wherein:a target domain enhanced user representation representing a cold boot user; />Representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Is defined by the target domain feature vector of (a); />Representing the trainable parameters.
According to the method, a plurality of key sets are sampled from the target domain user set by the cross-domain recommendation model of the cross-domain recommendation model, the attention score of each user in the key set and the target domain feature vector of the key set are calculated by combining the target domain fusion user representation and the attention mechanism, and then the target domain feature vector of each key set is fused to obtain the target domain enhancement user representation, so that knowledge of the target domain can be fully mined by combining a relational network, the defect that the existing cross-domain recommendation model ignores knowledge mining of the target domain is overcome, and further performance of the cross-domain recommendation model can be better improved by extracting proper target domain features, and accuracy of cross-domain commodity recommendation can be improved.
In the implementation process, the target domain end user representation of the cold start user is calculated through the following formula;
wherein:a target domain end user representation representing a cold boot user; />A target domain enhanced user representation representing a cold boot user; />Target domain representing cold-start userThe user representations are fused.
Calculating a target domain prediction score of the commodity to be recommended according to the following formula;
wherein:target domain end user representation representing cold boot user +.>Commodity representation with target domain to be recommendedIs the inner product of (i) cold start user u i For commodity v to be recommended j Target domain prediction scores of (2).
According to the method and the device, the target domain end user representation is obtained by fusing the target domain fusion user representation and the target domain enhancement user representation of the cold start user, so that the complementary knowledge migration and the target domain feature extraction can be effectively combined, and the comprehensiveness and the accuracy of cross-domain commodity recommendation can be improved.
In the implementation process, because the learning of the cross-domain recommendation model of the patent application is mainly based on public users in two domains, all parameters in the model are optimized based on monitoring signals of the public users. In order to learn the mapping function, most existing mapping function-based methods employ a mapping-oriented optimization process to learn the mapping function by directly minimizing the distance of the user's representation of the source domain transformed by the mapping function from the user's representation in the target domain.
However, because of the few interactive records of some users, the learned embedded representation of the user or commodity may be inaccurate, in order to reduce the impact of such inaccurate embedded representation, the present application remains consistent with the PTUPCDR, and a task-oriented optimization process is used to train the model, so that training of the model is directly directed at a recommended target, and the negative impact of the inaccuracy of the embedded representation can be reduced as much as possible.
When the cross-domain recommendation model is trained, calculating an optimization target through the following formula;
wherein:expressing an optimization target during cross-domain recommendation model training; />Representing a cold start user u i A real score set for a target domain of the commodity; />Representing a cold start user u i For commodity v to be recommended j Target domain prediction scores for (2); r is (r) ij Representing a cold start user u i For commodity v to be recommended j Target domain true score of (2).
In order to better illustrate the advantages of the technical scheme of the invention, the following experiment is disclosed in the embodiment.
1. Experimental setup
1.1, dataset
In order to be convenient for fair comparison with other models, the experiment adopts a data set adopted by most existing methods, namely an Amazon comment data set, which is a large-scale data set of commodity interaction behaviors of users in a real scene. Specifically, amazon-5cores dataset was used, and three popular categories of data were chosen from the 24 total categories: movie_and_tv (Movie), cds_and_vinyl (Music), and books (Book), and define three cross-domain task categories from the selected dataset, task one: movie-Music, task two: book-Movie and task three: book-Music. All data partitioning and processing details remain consistent with the existing PTUPCDR model. Detailed task data statistics are shown in table 1.
TABLE 1 statistics of different Cross-Domain task data sets information
1.2 evaluation index
The Amazon comment data set comprises the scores (0-5 scores) of each interactive behavior user on the commodity, and the scores of the predicted users on the commodity essentially belong to regression problems, so that the scores are consistent with the existing work, and the average absolute error (MAE) and the Root Mean Square Error (RMSE) are selected as evaluation indexes in the experiment, and are represented by the RMSE and the MAE for convenience. The specific calculation mode is as follows:
/>
wherein n represents the number of samples,representing model predictive value, y i Representing the true value.
1.3, baseline model
The jcttdr model (i.e., the cross-domain recommendation model) proposed by the present patent application can be generally classified as a mapping-based cross-domain recommendation algorithm, and does not resort to other auxiliary information besides user commodity interaction data, so that the model proposed by the present patent application is mainly compared with the same mapping-based cross-domain recommendation algorithm. Thus, the present experiment selected the following model as the baseline model for comparison:
(1) TGT. And a matrix decomposition model trained by using the target domain data is used.
(2) MF (from Obing Xie, zhijie Qia, jun Rao, et al International and Contextual Attention Network for Cold-start Multi-channel Matching in Recommendation). The CMF can be regarded as an extension of a matrix decomposition model MF, so that the CMF can be applied to a cross-domain recommendation system, and the embedded representation of a public user is consistent in a source domain and a target domain;
(3) EMCDR (from SeongKu Kang, junyuang Hwang, dongha Lee, et al Semi-Supervised Learning for Cross-Domain Recommendation to Cold-Start Users). CDRs are a very popular type of cross-domain recommendation algorithm, and many mapping-based cross-domain recommendation algorithms have been developed. Firstly, respectively learning user and commodity embedded representations of a source domain and a target domain through a hidden factor model, then learning a mapping function according to the learned embedded representations by using public users of the two domains, and finally, mapping the source domain user and the target domain by using the mapping function to recommend.
(4) DCDCSR (from Tong Man, huawei Shen, xiaolong Jin, et al Cross-Domain Recommendation: an Embedding and Mapping Approach). The CDCSR considers the influence of the sparsity of different users on the learned representation accuracy, so that the problem of inaccurate representation caused by excessive sparsity is relieved, and the robustness of the model is improved.
(5) SSCDR (from Fu, wenjing, zhaohui Peng, senzhang Wang, et al Deeply Fusing Reviews and Contents for Cold Start Users in Cross-Domain Recommendation Systems). SCDR takes into account the problem that mapping functions are prone to overfitting when the source and target domains overlap by too few users, and applies semi-supervised learning to alleviate this problem.
(6) PTUPCDR (from Zhu, yongchun, zhenwei Tang, yuman Liu et al Personalized Transfer of User Preferences for Cross-domain Recommendation). TUPCDR learns a mapping function for each user by using element learning technology, so that the model can fully consider the personalized characteristics of each user and obtain the current optimal result.
2. Overall experiment
The overall experimental results are shown in tables 2 and 3, with the optimal results being indicated by bolded scale, imp% representing the relative percent improvement of jcttdr model of the present patent application over the best baseline performance.
Table 2 shows the performance of the JCTCDR model of the present application in comparison with other baseline models in terms of Mean Absolute Error (MAE) evaluation index, and it can be seen from the table that the performance of JCTCDR model in terms of Mean Absolute Error (MAE) index exceeds that of some baseline models, wherein the effect improvement in task one and task three is obvious and the improvement in task two is relatively small. The reason may be that for the first task and the third task, because there are fewer overlapping users in the two domains, the mapping function may not be learned sufficiently, and the source domain knowledge cannot be transferred to the target domain accurately, that is, the source domain knowledge has limited help to the target domain, and the user perception target domain feature extraction module provided by the present application can effectively make up for the deficiency through knowledge mining of the target domain, so that the lifting result is larger; for the second task, as more overlapping users exist, the existing baseline model can learn a better mapping function, and the improvement effect of the method of the patent application is relatively weakened, so that the final improvement is relatively small.
Table 3 shows that the JCTCDR model of the present patent application achieves optimal performance in comparison with other baseline model performance in Root Mean Square Error (RMSE) evaluation index, similar to Table 2, with significant improvement in the effects in task one and task three and relatively less improvement in task two. The specific cause analysis is consistent with table 2.
TABLE 2 comparison of JCTCDR models with other models on Mean Absolute Error (MAE) evaluation index
TABLE 3 comparison of JCTCDR model with other models on Root Mean Square Error (RMSE) evaluation index
3. Ablation experiments
In order to verify the effectiveness of the personality-commonality complementary mapping module and the target domain feature extraction module perceived by the user, the present experiment further carries out the following ablation experiment, wherein w/o com represents that the personality-commonality complementary mapping module is replaced by a personality mapping function module proposed by the PTUPCDR model, namely, only the personality attribute of the user is considered, and the common attribute is ignored; w/o t _fea means that the user-aware target domain feature extraction module proposed in the present application is removed, i.e. the knowledge of source domain migration is only used without considering the mining and utilization of the target domain knowledge itself. The specific results are shown in tables 4 and 5.
According to tables 4 and 5, it can be found that the addition of any single module provided by the present patent application can obtain results better than all baseline models, especially the improvement of the personality-commonality complementary mapping module is more remarkable, while the improvement of the performance of the user-perceived target domain feature extraction module is relatively smaller, which indicates that the knowledge of the source domain is still the main source of the performance improvement, and the effect of the features extracted by the target domain is more in the compensation of the relatively insufficient knowledge of the source domain, which not only proves the effectiveness of the two modules provided by the present patent application, but also demonstrates the reliability of the guess provided in the overall experiment.
Notably, in a few cases, the two modules proposed in the present application will be slightly less effective when applied simultaneously than when only the personality-commonality complementary mapping module is applied, presumably because noise may also be introduced when introducing the target domain features, resulting in reduced model performance.
TABLE 4 ablation experiments on Mean Absolute Error (MAE)
Table 5 ablation experiments for Root Mean Square Error (RMSE)
4. General purpose experiment
To verify the generalization ability of the model proposed in the present patent application on different underlying models, the present experiment further demonstrates the performance comparison of the model proposed in the present patent application on MF and GMF as underlying models with other cross-domain recommended models, and the specific comparison results are shown in fig. 3, fig. 4, fig. 5 and fig. 6.
It can be known that the model proposed by the present application can obtain the optimal performance no matter the model uses MF as the bottom model or GMF as the bottom model, which indicates that the JCTCDR model proposed by the present application has a certain commonality.
5. Parameter sensitivity experiment
To verify the robustness of the model proposed in this patent application, the present experiment tested the influence of the number of key sets and the number of users contained in each key set on the model performance, it should be noted that when the influence of the number of key sets on the model performance is tested, the number of users in each key set is fixed to be 100, and similarly, when the influence of the number of users in the key set on the model performance is tested, the number of key sets sampled is fixed to be 4. In addition, in order to reflect the influence of the super-parameters as truly as possible, the experimental result of each task is based on β=50%, i.e., the training set and the test set each account for half. Specific experimental results are shown in fig. 7, 8, 9 and 10 below. According to experimental results, overall, the influence of different super-parameter settings on the final performance of the model is smaller, and the JCTCDR model is stable for the super-parameter settings. The combination of super-parameters at the time of achieving optimal performance is different for each task. Specifically, for task one, when the number of key sets is fixed to be 4, the best result is obtained when the number of users is 100; when the number of users is fixed to be 100, the value of the number of the key sets is 4, and the best result is obtained; similar results are generally obtained for task two; however, for task three, the results are somewhat different from the first two tasks, and good results are obtained when 150 or 250 users and 3 or 5 key sets are used respectively, but the overall result fluctuation is not large.
Finally, it should be noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the technical solution, and those skilled in the art should understand that modifications and equivalents may be made to the technical solution of the present invention without departing from the spirit and scope of the present invention, and all such modifications and equivalents are included in the scope of the claims.
Claims (8)
1. The cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction is characterized by comprising the following steps:
s1: acquiring a plurality of commodities to be recommended of a cold start user existing in a target domain in the target domain;
s2: inputting the source domain user representation of the cold start user into the trained cross-domain recommendation model, and outputting a corresponding target domain end user representation; the cross-domain recommendation model comprises a complementary mapping module and a feature extraction module;
the complementary mapping module is used for generating a corresponding target domain personality representation and a target domain commonality representation based on the source domain user representation of the cold start user; then fusing the corresponding target domain personality representation and target domain commonality representation to obtain a target domain fusion user representation of the cold start user;
generating a target domain personality representation of the cold start user by:
S201: commodity interaction sequence for cold start user in source domainInputting a source domain recommendation model, and outputting an embedded vector representation of the obtained commodity interaction sequence>n represents the number of interactive items;
wherein:indicating that the user is at t k The embedded vectors of the goods are interacted at each moment; />Indicating that the user is at t k Goods interacted at each moment; />Representing a source domain recommendation model;
s202: computing cold start user-to-commodity interaction sequence based on attention mechanismThe attention score of each commodity in the commodity interaction sequence is further represented by an embedded vector of the commodity interaction sequence>The embedded vectors of the commodities are combined with the corresponding attention scores to generate user characteristics based on the commodities;
wherein:representing commodity-based user characteristics; alpha j Representing a cold start user u i Commodity interaction sequence->Commodity v of China j Is the concentration score of (1), wherein +.>Attn (·) representsAttention network,/->Parameters representing an attention network; />Embedded vector representation representing a merchandise interaction sequence +.>An embedded vector of the commodity; a, a l Representing the attention score prior to normalization, i.e. the output of the attention network; />Representing a commodity;
s203: taking the commodity-based user characteristics as the input of the meta-network, and taking the output of the meta-network as the weight parameter of the personalized characteristic mapping model; then the source domain user representation of the cold start user is used as the input of the personality characteristic mapping model, and the output of the personality characteristic mapping model is used as the target domain personality representation of the cold start user;
Wherein:the output of the meta-network, namely the weight parameter of the personality characteristic mapping model; />Representing commodity-based user characteristics; f (f) meta (-) represents a meta-network; epsilon represents a meta-network parameter; />The output of the personality characteristic mapping model is represented, namely the personality of the target domain of the cold start user is represented; f (f) p Representing a personality trait map model; />Representing a cold start user u i Source domain user representation of (2);
the source domain user representation of the cold start user is used as input of a commonality feature mapping model, and the output of the commonality feature mapping model is used as target domain commonality representation of the cold start user;
wherein:the output of the common feature mapping model is represented, namely the target domain common representation of the cold start user; />Representing a cold start user u i Source domain user representation of (2); f (f) com (. Cndot.) represents a commonality feature mapping model; omega represents model parameters;
the feature extraction module is used for sampling a plurality of user sets from the target domain user sets to serve as key sets; then combining the target domain fusion user representations of the cold start users, and calculating the attention scores of the target domain fusion user representations corresponding to the cold start users for the users in the corresponding key set by using an attention mechanism; calculating target domain fusion user expression of the cold start user by combining the attention scores of the users in the corresponding key sets, and representing target domain feature vectors of the corresponding key sets; finally, fusing target domain fusion user expression of the cold start user with target domain feature vectors of each key set to obtain target domain enhancement user expression of the cold start user;
Fusing target domain fusion user representation and target domain enhancement user representation of a cold start user by a cross-domain recommendation model to obtain corresponding target domain end user representation;
s3: calculating a target domain prediction score of the corresponding cold start user for the commodity to be recommended based on the target domain end user representation of the cold start user in combination with the corresponding target domain commodity representation to be recommended;
s4: and taking one or more commodities to be recommended with highest target domain prediction scores as recommendation results of the cold start user.
2. The cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction of claim 1, wherein: in step S2, calculating a target domain fusion user representation of the cold start user through the following formula;
wherein:a target domain fusion user representation representing a cold boot user; />A target domain personality representation representing a cold boot user; />A target domain commonality representation representing a cold-start user.
3. The cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction of claim 1, wherein: in step S2, the attention score of the target domain fusion user representation of the cold start user for the corresponding user in the corresponding key set is calculated through the following formula;
Wherein:representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Middle user->Is a fraction of the attention of (2); />A target domain fusion user representation representing a cold boot user; all represent trainable parameters; />Representing a join operation; />Representing a vector representation corresponding to the user.
4. The cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction of claim 3, wherein: in step S2, calculating a target domain feature vector of a target domain fusion user representation of the cold start user for the key set through the following formula;
wherein:representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Is defined by the target domain feature vector of (a); />Representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Middle user->Is a fraction of the attention of (2); />Representing the trainable parameters.
5. The cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction of claim 4, wherein: in step S2, the target domain enhanced user representation of the cold start user is calculated by the following formula;
wherein:indicating cold startThe target domain of the user enhances the user representation; / >Representing a cold start user u i Is a target domain fusion user representation +.>For the kth z Key set->Is defined by the target domain feature vector of (a); />Representing the trainable parameters.
6. The cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction of claim 1, wherein: in step S2, calculating a target domain end user representation of the cold start user by the following formula;
wherein:a target domain end user representation representing a cold boot user; />A target domain enhanced user representation representing a cold boot user; />The target domain representing the cold-boot user fuses the user representations.
7. The cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction of claim 1, wherein: in step S3, calculating a target domain prediction score of the commodity to be recommended according to the following formula;
wherein:target domain end user representation representing cold boot user +.>Commodity representation +.>Is the inner product of (i) cold start user u i For commodity v to be recommended j Target domain prediction scores of (2).
8. The cross-domain recommendation method for combining complementary knowledge migration and target domain feature extraction of claim 1, wherein: in step S2, when the cross-domain recommendation model is trained, calculating an optimization target through the following formula;
Wherein:expressing an optimization target during cross-domain recommendation model training; />Representing a set of true scores of public users of the source domain and the target domain in the target domain; />Representing a cold start user u i For commodity v to be recommended j Target domain prediction scores for (2); r is (r) ij Indicating coldStart user u i For commodity v to be recommended j Target domain true score of (2).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211313565.6A CN116051224B (en) | 2022-10-25 | 2022-10-25 | Cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211313565.6A CN116051224B (en) | 2022-10-25 | 2022-10-25 | Cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116051224A CN116051224A (en) | 2023-05-02 |
| CN116051224B true CN116051224B (en) | 2024-01-23 |
Family
ID=86132049
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211313565.6A Active CN116051224B (en) | 2022-10-25 | 2022-10-25 | Cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116051224B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111291274A (en) * | 2020-03-02 | 2020-06-16 | 苏州大学 | Article recommendation method, device, equipment and computer-readable storage medium |
| CN112417298A (en) * | 2020-12-07 | 2021-02-26 | 中山大学 | Cross-domain recommendation method and system based on a small number of overlapped users |
| CN113435984A (en) * | 2021-08-27 | 2021-09-24 | 苏州浪潮智能科技有限公司 | Cross-domain recommendation method, system, storage medium and equipment |
| CN113761389A (en) * | 2021-08-18 | 2021-12-07 | 淮阴工学院 | Cross-domain recommendation method based on subject label |
| CN115098692A (en) * | 2022-05-30 | 2022-09-23 | 清华大学 | Cross-domain recommendation method and device, electronic equipment and storage medium |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10706358B2 (en) * | 2016-05-13 | 2020-07-07 | Cognitive Scale, Inc. | Lossless parsing when storing knowledge elements within a universal cognitive graph |
-
2022
- 2022-10-25 CN CN202211313565.6A patent/CN116051224B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111291274A (en) * | 2020-03-02 | 2020-06-16 | 苏州大学 | Article recommendation method, device, equipment and computer-readable storage medium |
| CN112417298A (en) * | 2020-12-07 | 2021-02-26 | 中山大学 | Cross-domain recommendation method and system based on a small number of overlapped users |
| CN113761389A (en) * | 2021-08-18 | 2021-12-07 | 淮阴工学院 | Cross-domain recommendation method based on subject label |
| CN113435984A (en) * | 2021-08-27 | 2021-09-24 | 苏州浪潮智能科技有限公司 | Cross-domain recommendation method, system, storage medium and equipment |
| CN115098692A (en) * | 2022-05-30 | 2022-09-23 | 清华大学 | Cross-domain recommendation method and device, electronic equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116051224A (en) | 2023-05-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111222332B (en) | Commodity recommendation method combining attention network and user emotion | |
| CN111797321B (en) | Personalized knowledge recommendation method and system for different scenes | |
| De Winter et al. | Combining temporal aspects of dynamic networks with node2vec for a more efficient dynamic link prediction | |
| CN108563755A (en) | A kind of personalized recommendation system and method based on bidirectional circulating neural network | |
| CN110874439B (en) | Recommendation method based on comment information | |
| CN110245285A (en) | A kind of personalized recommendation method based on Heterogeneous Information network | |
| CN107016122B (en) | Knowledge recommendation method based on time migration | |
| CN113761359B (en) | Data packet recommendation method, device, electronic equipment and storage medium | |
| CN113191838A (en) | Shopping recommendation method and system based on heterogeneous graph neural network | |
| CN113806630A (en) | Attention-based multi-view feature fusion cross-domain recommendation method and device | |
| CN115203550A (en) | Social recommendation method and system for enhancing neighbor relation | |
| CN113742586B (en) | Learning resource recommendation method and system based on knowledge graph embedding | |
| CN115098692A (en) | Cross-domain recommendation method and device, electronic equipment and storage medium | |
| CN110851694A (en) | Personalized recommendation system based on user memory network and tree structure depth model | |
| CN116051224B (en) | Cross-domain recommendation method combining complementary knowledge migration and target domain feature extraction | |
| Chen et al. | A multitask recommendation algorithm based on DeepFM and Graph Convolutional Network | |
| CN115935067A (en) | Article recommendation method integrating semantics and structural view for socialized recommendation | |
| CN106650972B (en) | Recommendation system score prediction method based on cloud model and oriented to social network | |
| CN115080856A (en) | Recommendation method and device and training method and device of recommendation model | |
| CN113327154B (en) | E-commerce user message pushing method and system based on big data | |
| CN115525819A (en) | Cross-domain recommendation method for information cocoon room | |
| CN114862514A (en) | User preference commodity recommendation method based on meta-learning | |
| CN110956528B (en) | Recommendation method and system for e-commerce platform | |
| Liu et al. | A Graph Neural Network Recommendation Method Integrating Multi Head Attention Mechanism and Improved Gated Recurrent Unit Algorithm | |
| Chang et al. | Construction of a personalised online learning resource recommendation model based on self-adaptation |
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 |