CN106651519B - Personalized recommendation method and system based on label information - Google Patents
Personalized recommendation method and system based on label information Download PDFInfo
- Publication number
- CN106651519B CN106651519B CN201611207428.9A CN201611207428A CN106651519B CN 106651519 B CN106651519 B CN 106651519B CN 201611207428 A CN201611207428 A CN 201611207428A CN 106651519 B CN106651519 B CN 106651519B
- Authority
- CN
- China
- Prior art keywords
- product
- matrix
- user
- label
- model
- 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 41
- 239000011159 matrix material Substances 0.000 claims abstract description 149
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 34
- 230000002452 interceptive effect Effects 0.000 claims abstract 4
- 230000003993 interaction Effects 0.000 claims description 55
- 239000013598 vector Substances 0.000 claims description 25
- 238000010276 construction Methods 0.000 claims description 3
- 238000012163 sequencing technique Methods 0.000 claims description 3
- 238000002372 labelling Methods 0.000 abstract 2
- 230000006870 function Effects 0.000 description 6
- 238000012360 testing method Methods 0.000 description 5
- 230000001186 cumulative effect Effects 0.000 description 4
- 230000004927 fusion Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000011478 gradient descent method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 208000003443 Unconsciousness Diseases 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 230000009466 transformation Effects 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]
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of personalized recommendation method and system based on label information, this method comprises: generating three-dimensional table to the state that each product identification is marked by each label information according to corresponding to each user identifier, consumer products Interactive matrix model and Product labelling relational matrix model are established according to three-dimensional table；According to consumer products Interactive matrix model and Product labelling relational matrix model, user identifier, product identification and the corresponding joint decomposition model of label information are constructed；Joint decomposition model, which is solved, using Bayes's personalized ordering method obtains multiple parameter values；User identifier is obtained to the preference of each product identification according to parameter value；Recommendation mark is chosen from product identification according to preference, will recommend to identify the terminal where corresponding product information recommends corresponding user identifier.In this way, can solve the limitation of the Deta sparseness in label information, the precision of personalized ordering is improved, to improve the accuracy of recommendation.
Description
Technical Field
The invention relates to the technical field of information, in particular to a personalized recommendation method and system based on label information.
Background
With the development of information technology, electronic commerce develops rapidly, and the phenomenon of information overload of e-commerce platforms is becoming more serious. To alleviate information overload, e-commerce platforms typically use personalized recommendation systems. The personalized recommendation system constructs a user interest preference model according to the online browsing data or purchasing data of the user individuals, so that products meeting the unique requirements of the users are recommended to the users, the user experience can be optimized, and the platform user flow can be improved.
The traditional personalized recommendation system recommends products with the same type of labels to users according to the labels marked on the purchased or browsed products by the users. However, users generally label products with few labels, and a product contains a limited number of labels, so that the e-commerce platform includes label information with sparsity, the label information cannot accurately represent descriptions of the products, and at the same time, due to the sparsity of the label information, the similarity between users is inaccurate, so that the recommendation accuracy is not high.
Disclosure of Invention
In view of the above, it is necessary to provide a tag information-based personalized recommendation method and system with recommendation accuracy.
A personalized recommendation method based on tag information comprises the following steps:
acquiring user identification, product identification and label information of the e-commerce platform, and marking states of the product identifications by the label information corresponding to the user identifications, and generating a three-dimensional table according to the acquired states;
establishing a user product interaction matrix model and a product label relation matrix model according to the three-dimensional table;
constructing a combined decomposition model corresponding to the user identification, the product identification and the label information according to the user product interaction matrix model and the product label relation matrix model;
solving the joint decomposition model by using a Bayes personalized sorting method to obtain a plurality of parameter values;
acquiring the preference degree of the user identification to each product identification according to the parameter value;
and sorting the product identifications according to the sequence of the preference degrees from large to small, sequentially selecting recommended identifications from the sorted product identifications, and recommending the product information corresponding to the recommended identifications to the terminal where the corresponding user identification is located.
A tag information-based personalized recommendation system, comprising:
the three-dimensional table acquisition module is used for acquiring user identification, product identification and label information of the e-commerce platform, and the marking state of each product identification by each label information corresponding to each user identification, and generating a three-dimensional table according to the acquired state;
the matrix model establishing module is used for establishing a user product interaction matrix model and a product label relation matrix model according to the three-dimensional table;
the model combination construction module is used for constructing a combined decomposition model corresponding to the user identification, the product identification and the label information according to the user product interaction matrix model and the product label relation matrix model;
the model solving module is used for solving the joint decomposition model by utilizing a Bayes personalized sorting method to obtain a plurality of parameter values;
the preference degree acquisition module is used for acquiring the preference degree of the user identifier to each product identifier according to the parameter value;
and the product recommending module is used for sequencing the product identifications according to the sequence of the preference degrees from large to small, sequentially selecting recommended identifications from the sequenced product identifications, and recommending the product information corresponding to the recommended identifications to the terminal where the corresponding user identification is located.
According to the tag information-based personalized recommendation method and system, a three-dimensional table is generated by marking the state of each product identifier according to each tag information corresponding to each user identifier, a user product interaction matrix model and a product tag relation matrix model are established according to the three-dimensional table, and a combined decomposition model corresponding to the user identifier, the product identifier and the tag information is established according to the user product interaction matrix model and the product tag relation matrix model; and then solving the joint decomposition model by using a Bayesian personalized sorting method to obtain a plurality of parameter values, obtaining the preference degree of the user identification to each product identification according to the parameter values, selecting a recommended identification from the product identifications according to the preference degree, and recommending the product information corresponding to the recommended identification to the terminal where the corresponding user identification is located. Therefore, by combining the tag information and the ternary relationship between the user identifier and the product identifier, a joint decomposition model is established, the joint decomposition model is learned by combining a Bayes personalized sorting method, personalized fusion tag sorting is performed, the limitation of data sparsity in the tag information can be solved, the precision of personalized sorting is improved, and the recommendation accuracy is improved.
Drawings
FIG. 1 is a flowchart illustrating a method for personalized recommendation based on tag information according to an embodiment;
FIG. 2 is a block diagram of a tag information-based personalized recommendation system in an embodiment.
Detailed Description
Referring to fig. 1, a personalized recommendation method based on tag information in an embodiment includes the following steps.
S110: and acquiring user identification, product identification and label information of the e-commerce platform, and marking states of the product identifications by the label information corresponding to the user identifications, and generating a three-dimensional table according to the acquired states.
The user identification is used for identifying a unique user, the product identification is used for identifying the type of a product, and the label information refers to information used by the user for marking the product. The state of the product identification marked by the label information corresponding to the user identification comprises marked state and unmarked state, and the ternary relationship among the user identification, the product identification and the label information in the three-dimensional table can be represented by numerical values of '1' and '0'. For example, a three-dimensional table may be functionally described as:
F＝(U,I,T,Y) (1)；
wherein, F represents a three-dimensional table, U identifies a user identification set, I represents a product identification set, T represents a label information set, Y represents a ternary relationship among the user identification, the product identification and the label information, and the three-dimensional array Y can be (Y ═ Y)u,i,t)∈R|U|×|I|×|T|And (4) showing. | U | represents the number of user identifiers, | I | represents the number of product identifiers, | T | represents the number of tag information, U represents the serial number of user identifiers, I represents the serial number of product identifiers, and T represents the serial number of tag information. If the user identification u specifies the tag information t to the product identification i, yu,i,t1, otherwise yu,i,t＝0。
S120: and establishing a user product interaction matrix model and a product label relation matrix model according to the three-dimensional table.
The user product interaction matrix model is used for reflecting the relationship between the user identification and the product identification; the product label relation matrix model is used for reflecting the relation between the product identification and the label information.
In one embodiment, step S120 includes steps (a1) to (a 3).
Step (a 1): and generating a user product matrix and a product label matrix according to the three-dimensional table.
The ternary relationship of equation (1) is decomposed into two-dimensional adjacency matrices AUIAnd AITRespectively representing a user product matrix and a product label matrix. If the user corresponding to the u-th user identifier marks the product corresponding to the i-th product identifier, thenOtherwiseSimilarly, if the product corresponding to the ith product identification is markedThe tag corresponding to the t-th tag information is obtainedOtherwiseUser product matrix AUIAnd product label matrix AITThe two-dimensional tables are shown in tables 1 and 2, respectively.
TABLE 1
TABLE 2
Step (a 2): and establishing a user product interaction matrix model according to the user product matrix.
Step (a 3): and establishing a product label relation matrix model according to the product label matrix.
In one embodiment, step (a3) includes steps (b1) through (b 4).
Step (b 1): and establishing a product co-occurrence matrix among the product identifications according to the product label matrix.
The rows and columns in the product co-occurrence matrix are both product identifiers, and the elements in the product co-occurrence matrix are the number of the same label information shared by the identified products in the corresponding rows and the corresponding columns. Therefore, it can be known from the product co-occurrence matrix whether the same tag information is shared between two product identifications, and the number of shared tag information.
Step (b 2): and acquiring point mutual information values among the product identifications according to the product co-occurrence matrix and the product label matrix.
Step (b 3): and acquiring an interaction value between corresponding product identifications according to the point mutual information value and a preset value.
Step (b 4): and establishing a product label relation matrix model according to the interaction value among the corresponding product identifications, the preset product hidden eigenvector and the preset context hidden eigenvector.
In this embodiment, the user product interaction matrix model is:
the step (b2) includes:
the step (b3) includes:
mij＝max{PMI(i,j)-logk,0} (4)；
the product label relationship matrix model is:
mij＝wi·cj T (5)；
wherein,representing the interaction value between the user identifier u and the product identifier i, mu representing the overall mean value of the data in the user product matrix, biRepresenting a deviation of the product identity i, buIdentifying a deviation, x, of u for the useruIdentify u preference for product identification for user, yiIdentifying the eigenvectors, inner products, of i for the productOverall preference of user identification u to product identification i; PMI (i, j) represents a point mutual information value between a product identifier i and a product identifier j, # (i, j) represents the number of labels of label information shared by the product identifier i and the product identifier j, # (i) represents the number of label information corresponding to the product identifier i in a product label matrix, # (j) represents the number of label information corresponding to the product identifier j in the product label matrix, | D | represents the number of nonzero elements in the product co-occurrence matrix; m isijRepresenting the interaction value between the product identification i and the product identification j, k is a preset value, and wiProduct latent feature vector representing product identification i, cjA context implicit feature vector representing a product identity j.
In the formula (4), k controls the data sparsity of the matrix and can be determined by a cross validation mode; if PMI (i, j) -logk > 0, then mijPMI (i, j) -logk; on the contrary, mij＝0。
S130: and constructing a combined decomposition model corresponding to the user identification, the product identification and the label information according to the user product interaction matrix model and the product label relation matrix model.
The joint decomposition model is formed by combining the user product interaction matrix model and the product label relation matrix model and is used for reflecting the relation among the user identification, the product identification and the label information.
In one embodiment, the joint decomposition model is:
wherein, buiIn representing a joint decomposition modelThe observation reliability of (2) is high,representing the value of the interaction, x, between the user identity u and the product identity iuPreference of user identification u to product identification i, yiIdentify the feature vector of i, m, for the productijRepresenting the value of the interaction between the product identity i and the product identity j, cjA context implicit feature vector representing a product identity j.
Assume that the characteristic vector y of the product identifier i in equation (2)iAnd the product implicit feature vector w of the product identifier i in the formula (5)iIs the same, and thus using equation (2) and equation (5), equation (6) can be constructed. Wherein, Regularization represents Regularization terms among user identification, product identification and context implicit feature vectors.
S140: and solving the joint decomposition model by using a Bayes personalized sorting method to obtain a plurality of parameter values.
The main idea of the bayesian personalized ranking model is to define the positive feedback (interest) as 1, whereas the negative feedback (not interest) or the unobserved feedback (unconscious) is 0. And the user identifier u, the product identifier i belonging to the positive feedback has higher rank than the product identifier j fed back negatively or not observed. The solving process of step S140 is as follows.
Step (c 1): defining the optimal parameter of the u-th user identifier for the given user identifier u, and obtaining the learning target of the joint decomposition model according to the Bayes personalized sorting method, namely, maximizing the posterior probability in the following formula (7)This problem can be formulated with a maximum posterior probability likelihood as follows:
in the formula, Θ represents a parameter set { X in the joint decomposition modelu,d,Yi,d,Yj,d,Ct,d,Cl,d,δUI,δICAnd (c) adding, in parallel,and the preference ranking of the user corresponding to the u-th user identifier to the products corresponding to all the product identifiers.Representing a preference likelihood function of the u-th user identifier obtained by the sample set to all product identifiers; p (Θ) represents a priori knowledge of the parameters. Wherein the likelihood functionCan be expressed as the following equation (8):
in the formula,and in the preference sorting of the user corresponding to the u-th user identifier to the products corresponding to all the product identifiers, the product identifier i is arranged in front of the product identifier j, and the sorting of each pair of product identifiers is independent of the sorting of other pairs of product identifiers.The preference probability of the u-th user identifier compared with the product identifier j to the product identifier i is expressed as the following formula (9):
in the formula,the difference between the individual preference of the user representing the u-th user identifier for the product of the i-th product identifier and the individual preference for the product corresponding to the j-th product identifier is the Indicating the degree of preference of the user identity u for each product identity i,representing the preference degree of the user identification u to each product identification j;to representSigmoid Logistic function of (1). According to the transformation, abbreviations are usedInstead of the formerAnd assuming that the prior knowledge p (theta) of the parameters obeys a mean value of 0 and the covariance matrix is sigmaΘIs high
A gaussian distribution, the logarithmic form of the posterior probability in equation (7) can be simplified to equation (10) as follows:
therefore, in order to learn the user hidden feature matrix X and the product hidden feature matrix Y, the learning of the optimization objective function can be changed from the learning of equation (10) to the learning of equation (11) below:
similarly, for the learning of the product co-occurrence matrix based on the label information, the learning of the optimized objective function of the product implicit feature matrix Y and the upper and lower implicit feature matrices C can be constructed as the following formula (12):
where t and l represent different tag information, respectively. In conjunction with equations (11) and (12), learning the objective function of the joint decomposition model can be constructed as the following equation (13):
in formula (13), λΘIs the regularization coefficient of the model parameters and alpha is the parameter used to balance the two objective functions.
Step (c 2): and randomly initializing to generate a user hidden feature matrix X, a product hidden feature matrix Y and an upper and lower hidden feature matrix C.
Step (c 3): for the product co-occurrence matrix, calculating the product corresponding to the ith product identification by the user corresponding to the u user identificationIs different from the preference for the jth product identificationFurther obtain the parameter deltaUIValue of
Step (c 4): parameter X in parameter set theta is obtained by using random gradient descent methodu,d、Yi,d、Yj,dAt parameter Yi,dAnd Yj,dFor constant value, for parameter deltaUIIteratively updating until convergence is reached, thereby obtaining a parameter Xu,dWhile using the obtained parameter Xu,dFor parameter deltaUIIteratively updating until convergence is reached, thereby obtaining a parameter Yi,dAnd Yj,dThe iterative model parameter modification process is as follows:
Xu,d←Xu,d+μ(α·δUI(Yi,d-Yj,d)-λ·Xu,d)；
Yi,d←Yi,d+μ(α·δUI·Xu,d-λ·Yi,d)；
Yj,d←Yj,d+μ(α·δUI·(-Xu,d)-λ·Yj,d)；
in the formula, d represents the number of iterations, Xu,dPreference of user identification u for product identification, Y, in the d-th iterationi,dIdentifying the feature vector of i, Y, for the product in the d-th iterationj,dIdentifying the characteristic vector of j for the product in the d iteration, wherein alpha and lambda are regularization term coefficients, and mu is a learning rate;
step (c 5): for a product co-occurrence matrix based on the label information, calculating the difference value of the preference of the ith product identifier to the tth label information and the preference of the ith product identifier to the ith label informationFurther obtain the parameter deltaICHas a value of
Step (c 6): parameter set theta parameter Y is obtained by utilizing random gradient descent methodi,d、Ct,d、Cl,dAt parameter Ct,dAnd Cl,dFor constant value, for parameter deltaICIteratively updating until convergence is reached, thereby obtaining a parameter Yi,dThe optimum value of (d); at the same time, using the obtained parameter Yi,dFor parameter deltaICIteratively updating until convergence is reached, thereby obtaining a parameter Ct,dAnd Cl,dThe optimum value of (c). The model parameter iterative correction process is as follows:
Yi,d←Yi,d+μ((1-α)·δIC(Ct,d-Cl,d)-λ·Yi,d)；
Ct,d←Ct,d+μ((1-α)·δIC·Yi,d-λ·Ct,d)；
Cl,d←Cl,d+μ((1-α)·δIC·(-Yi,d)-λ·Cl,d)；
in the formula, Ct,dA context implicit feature vector, C, represented as a product identifier i marked by the t-th label information in the d-th iterationl,dAnd identifying the context hidden feature vector of the product i marked by the ith label information in the d iteration.
S150: and acquiring the preference degree of the user identification to each product identification according to the parameter value.
In one embodiment, step S150 includes:
wherein,indicating the degree of preference of the user identity u for the product identity i based on the tag information.
S160: and sorting the product identifications according to the sequence of the preference degrees from large to small, sequentially selecting recommended identifications from the sorted product identifications, and recommending the product information corresponding to the recommended identifications to the terminal where the corresponding user identification is located.
According to the tag information-based personalized recommendation method, a three-dimensional table is generated by marking the state of each product identifier according to each tag information corresponding to each user identifier, a user product interaction matrix model and a product tag relation matrix model are established according to the three-dimensional table, and a combined decomposition model corresponding to the user identifier, the product identifier and the tag information is established according to the user product interaction matrix model and the product tag relation matrix model; and then solving the joint decomposition model by using a Bayesian personalized sorting method to obtain a plurality of parameter values, obtaining the preference degree of the user identification to each product identification according to the parameter values, selecting a recommended identification from the product identifications according to the preference degree, and recommending the product information corresponding to the recommended identification to the terminal where the corresponding user identification is located. Therefore, by combining the tag information and the ternary relationship between the user identifier and the product identifier, a joint decomposition model is established, the joint decomposition model is learned by combining a Bayes personalized sorting method, personalized fusion tag sorting is performed, the limitation of data sparsity in the tag information can be solved, the precision of personalized sorting is improved, and the recommendation accuracy is improved.
The personalized recommendation method based on the label information can be used for personalized sequencing systems of entity products such as clothes and mobile phones, digital products such as movies and music, service products such as travel routes and online detection arrangement, and is wide in application range.
Referring to fig. 2, the tag information-based personalized recommendation system in an embodiment includes a three-dimensional table obtaining module 110, a matrix model building module 120, a model combination building module 130, a model solving module 140, a preference degree obtaining module 150, and a product recommendation module 160.
The three-dimensional table obtaining module 110 is configured to obtain a user identifier, a product identifier, and tag information of the provider platform, and a state of each product identifier marked by each tag information corresponding to each user identifier, and generate a three-dimensional table according to the obtained state.
The matrix model building module 120 is configured to build a user product interaction matrix model and a product label relationship matrix model according to the three-dimensional table.
The model combination construction module 130 is configured to construct a joint decomposition model corresponding to the user identifier, the product identifier, and the tag information according to the user product interaction matrix model and the product tag relationship matrix model.
The model solving module 140 is configured to solve the joint decomposition model by using a bayesian personalized ranking method to obtain a plurality of parameter values.
The preference degree obtaining module 150 is configured to obtain a preference degree of the user identifier for each product identifier according to the parameter value.
The product recommending module 160 is configured to sort the product identifiers according to a sequence of preference degrees from large to small, sequentially select recommended identifiers from the sorted product identifiers, and recommend the product information corresponding to the recommended identifiers to the terminal where the corresponding user identifier is located.
In one embodiment, the matrix model building module 120 includes a matrix generating unit (not shown), a first model building unit (not shown), and a second model building unit (not shown).
The matrix generating unit is used for generating a user product matrix and a product label matrix according to the three-dimensional table. The first model establishing unit is used for establishing a user product interaction matrix model according to the user product matrix. The second model establishing unit is used for establishing a product label relation matrix model according to the product label matrix.
In an embodiment, the second model building unit includes a product co-occurrence matrix generating subunit (not shown), a point-to-point mutual information value obtaining subunit (not shown), an interaction value obtaining subunit (not shown), and a model obtaining subunit (not shown).
And the product co-occurrence matrix generation subunit is used for establishing a product co-occurrence matrix among the product identifications according to the product label matrix. And the point mutual information value acquisition subunit is used for acquiring point mutual information values among the product identifications according to the product co-occurrence matrix and the product label matrix. And the interaction value acquisition subunit is used for acquiring an interaction value between corresponding product identifications according to the point mutual information value and the preset value. The model obtaining subunit is used for establishing a product label relation matrix model according to the interaction values among the corresponding product identifications, the preset product hidden feature vectors and the preset context hidden feature vectors.
In one embodiment, the user-product interaction matrix model is:
the point-to-point information value obtaining subunit is specifically configured to calculate:
the interaction value obtaining subunit is specifically configured to calculate:
mij＝max{PMI(i,j)-logk,0}；
the product label relationship matrix model is:
mij＝wi·cj T；
wherein,representing the interaction value between the user identifier u and the product identifier i, mu representing the overall mean value of the data in the user product matrix, biRepresenting a deviation of the product identity i, buIdentifying a deviation, x, of u for the useruIdentify u preference for product identification for user, yiIdentifying the eigenvectors, inner products, of i for the productOverall preference of user identification u to product identification i; PMI (i, j) represents a point mutual information value between a product identifier i and a product identifier j, # (i, j) represents the number of labels of label information shared by the product identifier i and the product identifier j, # (i) represents the number of label information corresponding to the product identifier i in a product label matrix, # (j) represents the number of label information corresponding to the product identifier j in the product label matrix, | D | represents a non-co-occurrence matrix in the product co-occurrence matrixThe number of zero elements; m isijRepresenting the interaction value between the product identification i and the product identification j, k is a preset value, and wiProduct latent feature vector representing product identification i, cjA context implicit feature vector representing a product identity j.
In an embodiment, the preference level obtaining module 150 is configured to obtain the preference level according to:
and calculating the preference degree of the user identification to each product identification. Wherein,indicating the degree of preference of the user identity u for the product identity i based on the tag information.
In the tag information-based personalized recommendation system, a three-dimensional table is generated by the three-dimensional table obtaining module 110 according to the state of each product identifier marked by each tag information corresponding to each user identifier, the matrix model building module 120 builds a user product interaction matrix model and a product tag relation matrix model according to the three-dimensional table, and the model combination building module 130 builds a joint decomposition model corresponding to the user identifier, the product identifier and the tag information according to the user product interaction matrix model and the product tag relation matrix model; the model solving module 140 solves the joint decomposition model by using a bayesian personalized ranking method to obtain a plurality of parameter values, the preference degree obtaining module 150 obtains the preference degree of the user identifier to each product identifier according to the parameter values, and the product recommending module 160 selects a recommended identifier from the product identifiers according to the preference degree and recommends the product information corresponding to the recommended identifier to the terminal where the corresponding user identifier is located. Therefore, a combined decomposition model is established by combining the tag information and the ternary relationship between the user identification and the product identification, the combined decomposition model is learned by combining the Bayes personalized sorting method, a novel personalized fusion tag sorting algorithm is provided, the limitation of data sparsity in the tag information can be solved, the precision of personalized sorting is improved, and the recommendation accuracy is improved.
Using the standard data set and the basic algorithm to learn relevant performance indexes on the machine, such as: comparing the area under the curve (AUC), the average accuracy (MAP), the normalized cumulative discount rate (NDCG), the average reverse rank (MRR) and the like, the superiority of the personalized recommendation method based on the label information can be verified.
The following experimental demonstration is carried out on the tag information-based personalized recommendation method in combination with a specific application case, and specifically includes:
1) preparing a standard data set
The method uses the last. Fm data set is a personalized recommendation data set published on the second international seminar of information heterogeneity and convergence of recommendation systems. Fm data set, where the user tags the music album or artist, including artist information, user-friendly relationship information, and tag information, includes 1737 independent users, 5632 pieces of music, and 14192 pieces of tag information. The training set and the test set are partitioned by using a rule of 80%/20%.
2) Evaluation index
The area under the curve (AUC), the average accuracy (MAP), the recall ratio of the recommended length N (Rec @ N), the normalized cumulative discount ratio (NDCG), and the reciprocal average rank (MRR) were used as evaluation indexes in the present embodiment. The area under the curve AUC is used for evaluating the quality of a binary classifier; the Rec @ N evaluates the ability of a group recommendation system to return all relevant products; the average accuracy MAP is used for predicting the average score of products, the normalized cumulative discount rate NDCG is used for obtaining the sorting advantages of related products, and the sorting effect of the average sorting reciprocal MRR inspection method is used for reflecting the accuracy of the related products in a recommendation list. The recall rate Rec @ N with the length of the recommendation list N is calculated by the formula:
n is the number of related products in the test set; n is a radical ofrelatedIs in rowSequence charts and number of products in the test set occurring simultaneously. The average accuracy mean MAP is defined as:
where | U | is the number of users in the test set,is an indicator variable, if the product ranked at the nth in the recommendation list of the user u also appears in the test set, the value is 1, and the other cases are all 0. Prec @ n denotes the accuracy of the recommendation until the product ranked as n is reached:
wherein n isrelatedIs the number of related products in the ordered list. The average inverse MRR is the inverse of the first correctly ordered product, and is calculated as:
the calculation of the normalized cumulative discount rate NDCG is relatively complex, and the calculation formula is:
3) experiments were performed on standard data sets
In order to verify the effectiveness of the personalized recommendation method based on the tag information, modeling and prediction are respectively carried out on data set distribution lengths of 5 and 10 of last. The results of the experiment are shown in table 3.
TABLE 3
Compared with the prediction results of recommendation models such as Random, MostPop, UserKNN, ItemKNN, IMF and BPRMF, the prediction results of the personalized recommendation method (Tag2Vec) based on the label information show better prediction accuracy on most indexes.
The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (10)
1. A personalized recommendation method based on tag information is characterized by comprising the following steps:
acquiring user identification, product identification and label information of an e-commerce platform, and marking states of the product identifications by the label information corresponding to the user identifications, and generating a three-dimensional table according to the user identifications, the product identifications, the label information and the acquired states;
establishing a user product interaction matrix model and a product label relation matrix model according to the three-dimensional table;
constructing a combined decomposition model corresponding to the user identification, the product identification and the label information according to the user product interaction matrix model and the product label relation matrix model;
solving the joint decomposition model by using a Bayes personalized sorting method to obtain a plurality of parameter values;
acquiring the preference degree of the user identification to each product identification according to the parameter value;
and sorting the product identifications according to the sequence of the preference degrees from large to small, sequentially selecting recommended identifications from the sorted product identifications, and recommending the product information corresponding to the recommended identifications to the terminal where the corresponding user identification is located.
2. The tag information-based personalized recommendation method according to claim 1, wherein the establishing a user product interaction matrix model and a product tag relationship matrix model according to the three-dimensional table comprises:
generating a user product matrix and a product label matrix according to the three-dimensional table;
establishing the user product interaction matrix model according to the user product matrix;
and establishing the product label relation matrix model according to the product label matrix.
3. The tag information-based personalized recommendation method according to claim 2, wherein the establishing of the product tag relation matrix model according to the product tag matrix comprises:
establishing a product co-occurrence matrix among product identifications according to the product label matrix;
acquiring point mutual information values among the product identifications according to the product co-occurrence matrix and the product label matrix;
acquiring an interaction value between corresponding product identifications according to the point interaction information value and a preset value;
and establishing the product label relation matrix model according to the interaction value among the corresponding product identifications, the preset product hidden feature vector and the preset context hidden feature vector.
4. The tag information-based personalized recommendation method according to claim 3, wherein the user product interaction matrix model is:
the obtaining of the mutual point information values between the product identifications according to the product co-occurrence matrix and the product label matrix includes:
the obtaining of the interaction value between corresponding product identifiers according to the point-to-point interaction information value and the preset value includes:
mij＝max{PMI(i,j)-logk,0}；
the product label relation matrix model is as follows:
mij＝wi·cj T；
wherein,representing the interaction value between a user identifier u and a product identifier i, mu representing the overall mean value of data in the user product matrix, biRepresenting a deviation of the product identity i, buIdentifying a deviation, x, of u for the useruIdentify u preference for product identification for user, yiIdentifying the eigenvectors, inner products, of i for the productOverall preference of user identification u to product identification i; PMI (i, j) represents a mutual point information value between a product identifier i and a product identifier j, # (i, j) represents the number of labels of label information shared by the product identifier i and the product identifier j, # (i) represents the number of label information corresponding to the product identifier i in the product label matrix, # (j) represents the number of label information corresponding to the product identifier j in the product label matrix, | D | represents the number of nonzero elements in the product co-occurrence matrix; m isijRepresenting the interaction value between the product identification i and the product identification j, k being the preset value, wiProduct latent feature vector representing product identification i, cjA context implicit feature vector representing a product identity j.
5. The tag information-based personalized recommendation method according to claim 1, wherein the joint decomposition model is:
wherein, buiRepresenting in the joint decomposition modelThe observation reliability of (2) is high,representing the value of the interaction, x, between the user identity u and the product identity iuPreference of user identification u to product identification i, yiIdentify the feature vector of i, m, for the productijRepresenting the value of the interaction between the product identity i and the product identity j, cjA context implicit feature vector representing a product identity j.
6. The tag information-based personalized recommendation method according to claim 1, wherein the obtaining of the preference degree of the user identifier for each product identifier according to the parameter value comprises:
wherein,indicating the degree of preference of the user identity u for the product identity i based on the tag information.
7. A personalized recommendation system based on tag information is characterized by comprising:
the three-dimensional table acquisition module is used for acquiring user identification, product identification and label information of the e-commerce platform, and the marking state of each product identification by each label information corresponding to each user identification, and generating a three-dimensional table according to the user identification, the product identification, the label information and the acquired state;
the matrix model establishing module is used for establishing a user product interaction matrix model and a product label relation matrix model according to the three-dimensional table;
the model combination construction module is used for constructing a combined decomposition model corresponding to the user identification, the product identification and the label information according to the user product interaction matrix model and the product label relation matrix model;
the model solving module is used for solving the joint decomposition model by utilizing a Bayes personalized sorting method to obtain a plurality of parameter values;
the preference degree acquisition module is used for acquiring the preference degree of the user identifier to each product identifier according to the parameter value;
and the product recommending module is used for sequencing the product identifications according to the sequence of the preference degrees from large to small, sequentially selecting recommended identifications from the sequenced product identifications, and recommending the product information corresponding to the recommended identifications to the terminal where the corresponding user identification is located.
8. The tag information-based personalized recommendation system according to claim 7, wherein the matrix model building module comprises:
the matrix generating unit is used for generating a user product matrix and a product label matrix according to the three-dimensional table;
the first model establishing unit is used for establishing the user product interaction matrix model according to the user product matrix;
and the second model establishing unit is used for establishing the product label relation matrix model according to the product label matrix.
9. The tag information-based personalized recommendation system according to claim 8, wherein the second model building unit comprises:
the product co-occurrence matrix generation subunit is used for establishing a product co-occurrence matrix among the product identifications according to the product label matrix;
a point mutual information value obtaining subunit, configured to obtain, according to the product co-occurrence matrix and the product label matrix, a point mutual information value between the product identifiers;
the interactive value acquisition subunit is used for acquiring an interactive value between corresponding product identifications according to the point mutual information value and a preset value;
and the model obtaining subunit is used for establishing the product label relation matrix model according to the interaction value among the corresponding product identifications, the preset product hidden feature vector and the preset context hidden feature vector.
10. The tag information-based personalized recommendation system according to claim 9, wherein the user product interaction matrix model is:
the point mutual information value obtaining subunit is specifically configured to calculate:
the interaction value obtaining subunit is specifically configured to calculate:
mij＝max{PMI(i,j)-logk,0}；
the product label relation matrix model is as follows:
mij＝wi·cj T；
wherein,representing the interaction value between a user identifier u and a product identifier i, mu representing the overall mean value of data in the user product matrix, biRepresenting a deviation of the product identity i, buIdentifying a deviation, x, of u for the useruIdentify u preference for product identification for user, yiIdentifying the eigenvectors, inner products, of i for the productOverall preference of user identification u to product identification i; PMI (i, j) represents a mutual point information value between a product identifier i and a product identifier j, # (i, j) represents the number of labels of label information shared by the product identifier i and the product identifier j, # (i) represents the number of label information corresponding to the product identifier i in the product label matrix, # (j) represents the number of label information corresponding to the product identifier j in the product label matrix, | D | represents the number of nonzero elements in the product co-occurrence matrix; m isijRepresenting the interaction value between the product identification i and the product identification j, k being the preset value, wiProduct latent feature vector representing product identification i, cjA context implicit feature vector representing a product identity j.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611207428.9A CN106651519B (en) | 2016-12-23 | 2016-12-23 | Personalized recommendation method and system based on label information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611207428.9A CN106651519B (en) | 2016-12-23 | 2016-12-23 | Personalized recommendation method and system based on label information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106651519A CN106651519A (en) | 2017-05-10 |
| CN106651519B true CN106651519B (en) | 2019-11-12 |
Family
ID=58826862
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611207428.9A Active CN106651519B (en) | 2016-12-23 | 2016-12-23 | Personalized recommendation method and system based on label information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106651519B (en) |
Families Citing this family (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107239993B (en) * | 2017-05-24 | 2020-11-24 | 海南大学 | Matrix decomposition recommendation method and system based on expansion label |
| CN107506398B (en) * | 2017-08-02 | 2020-01-24 | 杭州东信北邮信息技术有限公司 | Method for adding label attribute to book |
| CN107909498B (en) * | 2017-10-26 | 2020-07-28 | 厦门理工学院 | Recommendation method based on area below maximized receiver operation characteristic curve |
| CN107704631B (en) * | 2017-10-30 | 2020-12-01 | 西华大学 | Crowdsourcing-based music annotation atom library construction method |
| CN107833117B (en) * | 2017-12-13 | 2020-03-17 | 合肥工业大学 | Bayesian personalized sorting recommendation method considering tag information |
| CN108416628B (en) * | 2018-03-15 | 2020-07-07 | 重庆大学 | Restaurant dish intelligent recommendation system integrating food multi-attribute relationship |
| CN108763251B (en) * | 2018-04-02 | 2021-06-01 | 创新先进技术有限公司 | Personalized recommendation method and device for nuclear product and electronic equipment |
| CN108829808B (en) * | 2018-06-07 | 2021-07-13 | 麒麟合盛网络技术股份有限公司 | Page personalized sorting method and device and electronic equipment |
| CN110619447B (en) * | 2018-06-20 | 2023-03-24 | 广州虎牙信息科技有限公司 | Anchor evaluation method, device, equipment and storage medium |
| CN109034389A (en) * | 2018-08-02 | 2018-12-18 | 黄晓鸣 | Man-machine interactive modification method, device, equipment and the medium of information recommendation system |
| CN109087178B (en) * | 2018-08-28 | 2021-05-18 | 清华大学 | Commodity recommendation method and device |
| CN109272390A (en) * | 2018-10-08 | 2019-01-25 | 中山大学 | The personalized recommendation method of fusion scoring and label information |
| CN110008411B (en) * | 2019-04-17 | 2021-01-26 | 重庆大学 | Deep learning interest point recommendation method based on user sign-in sparse matrix |
| CN110674020B (en) * | 2019-09-02 | 2022-07-22 | 平安科技（深圳）有限公司 | APP intelligent recommendation method and device and computer readable storage medium |
| CN111177569B (en) * | 2020-01-07 | 2021-06-11 | 腾讯科技（深圳）有限公司 | Recommendation processing method, device and equipment based on artificial intelligence |
| CN111582991B (en) * | 2020-05-13 | 2023-09-01 | 中国银行股份有限公司 | Product information recommendation method and device |
| CN111861759B (en) * | 2020-06-15 | 2022-12-23 | 北京百分点科技集团股份有限公司 | Matching method and system for product and customer group |
| CN112487292A (en) * | 2020-11-27 | 2021-03-12 | 杭州闪援车管家汽车服务有限公司 | Vehicle maintenance system and method based on active recommendation |
| CN112801763B (en) * | 2021-04-14 | 2021-08-24 | 浙江口碑网络技术有限公司 | Touch and reach scheme generation method and device and electronic equipment |
| CN118096324B (en) * | 2024-04-23 | 2024-07-19 | 成都帆点创想科技有限公司 | Recommendation model training method and device |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103246672A (en) * | 2012-02-09 | 2013-08-14 | 中国科学技术大学 | Method and device for performing personalized recommendation on users |
| CN103425799A (en) * | 2013-09-04 | 2013-12-04 | 北京邮电大学 | Personalized research direction recommending system and method based on themes |
| CN105069666A (en) * | 2015-09-14 | 2015-11-18 | 浙江工商大学 | E-commerce personalized recommendation method integrated with user implicit information |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TW201118778A (en) * | 2009-11-17 | 2011-06-01 | Inst Information Industry | System and method for recommending product and automatic service equipment thereof |
-
2016
- 2016-12-23 CN CN201611207428.9A patent/CN106651519B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103246672A (en) * | 2012-02-09 | 2013-08-14 | 中国科学技术大学 | Method and device for performing personalized recommendation on users |
| CN103425799A (en) * | 2013-09-04 | 2013-12-04 | 北京邮电大学 | Personalized research direction recommending system and method based on themes |
| CN105069666A (en) * | 2015-09-14 | 2015-11-18 | 浙江工商大学 | E-commerce personalized recommendation method integrated with user implicit information |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106651519A (en) | 2017-05-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106651519B (en) | Personalized recommendation method and system based on label information | |
| CN108960719B (en) | Method and device for selecting products and computer readable storage medium | |
| CN110458627B (en) | Commodity sequence personalized recommendation method for dynamic preference of user | |
| CN105740401B (en) | A kind of interested site recommended method and device based on individual behavior and group interest | |
| CN106022865A (en) | Goods recommendation method based on scores and user behaviors | |
| CN107833117B (en) | Bayesian personalized sorting recommendation method considering tag information | |
| CN111191732A (en) | Target detection method based on full-automatic learning | |
| CN111062775A (en) | Recommendation system recall method based on attention mechanism | |
| CN106934035B (en) | Concept drift detection method in a kind of multi-tag data flow based on class and feature distribution | |
| CN104504055A (en) | Commodity similarity calculation method and commodity recommending system based on image similarity | |
| CN105740327B (en) | A kind of adaptively sampled method based on user preference | |
| CN103824213A (en) | Individualized recommendation method based on user preferences and commodity properties | |
| CN112115377A (en) | Graph neural network link prediction recommendation method based on social relationship | |
| CN110321473B (en) | Multi-modal attention-based diversity preference information pushing method, system, medium and device | |
| US8793201B1 (en) | System and method for seeding rule-based machine learning models | |
| CN106157156A (en) | A kind of cooperation recommending system based on communities of users | |
| CN107451210B (en) | Graph matching query method based on query relaxation result enhancement | |
| Singhal et al. | Towards a unified framework for visual compatibility prediction | |
| CN111340071A (en) | System and method for personalized product recommendation using hierarchical bayes | |
| CN105069129A (en) | Self-adaptive multi-label prediction method | |
| CN108920647B (en) | Low-rank matrix filling TOP-N recommendation method based on spectral clustering | |
| CN106127506A (en) | A kind of recommendation method solving commodity cold start-up problem based on Active Learning | |
| CN112632373A (en) | Personalized recommendation method and device based on probability matrix decomposition and computer readable storage medium | |
| CN110532429A (en) | It is a kind of based on cluster and correlation rule line on user group's classification method and device | |
| Liang et al. | Top-N-rank: A scalable list-wise ranking method for recommender systems |
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 |