CN116701861A - Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics - Google Patents

Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics Download PDF

Info

Publication number
CN116701861A
CN116701861A CN202310619062.XA CN202310619062A CN116701861A CN 116701861 A CN116701861 A CN 116701861A CN 202310619062 A CN202310619062 A CN 202310619062A CN 116701861 A CN116701861 A CN 116701861A
Authority
CN
China
Prior art keywords
user
model
item
ibpr
explicit
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310619062.XA
Other languages
Chinese (zh)
Inventor
冯军美
苗启广
韩玖胜
郗岳
黄婷
牛冠冲
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Guangzhou Institute of Technology of Xidian University
Original Assignee
Xidian University
Guangzhou Institute of Technology of Xidian University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xidian University, Guangzhou Institute of Technology of Xidian University filed Critical Xidian University
Priority to CN202310619062.XA priority Critical patent/CN116701861A/en
Publication of CN116701861A publication Critical patent/CN116701861A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to a post-fusion personalized recommendation model and a method based on explicit and implicit feedback characteristics, wherein the model comprises an explicit characteristic extraction module, an implicit characteristic extraction module and an integral characteristic extraction module, and the explicit characteristic extraction module and the implicit characteristic extraction module are respectively connected with the integral characteristic extraction module; the method integrates an IBPR model and a BiasSVD model, performs weighted summation on a predicted scoring matrix obtained by the BiasSVD model and a predicted ranking scoring matrix obtained by the IBPR model to obtain a final predicted ranking scoring matrix, ranks all ranking scores obtained by the prediction of a user from high to low, and recommends the previous item with the ranking being earlier to the user. According to the invention, the BiasSVD model is utilized to extract the explicit feedback characteristic, the IBPR model is utilized to extract the implicit feedback characteristic, and the historical scoring data and the implicit feedback data in the dataset are fully utilized, so that the cold start problem of the recommendation system is relieved, and the performance of the recommendation system is improved.

Description

Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics
Technical Field
The invention belongs to the technical fields of recommendation systems, information retrieval and data mining, and relates to a post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics.
Background
With explosive growth of data, recommendation systems are used to solve information overload problems. In the recommendation system, the recommendation algorithm is the most core and key part, and largely determines the performance of the recommendation system. Currently, there are many different recommendation algorithms, the most widely used recommendation algorithm being the collaborative filtering recommendation algorithm.
Collaborative filtering recommendation algorithms can be divided into two categories depending on the input data: collaborative filtering recommendation algorithm based on historical scores and collaborative filtering algorithm based on ranking scores. The input data of the former is explicit feedback data, such as historical scoring data of a user, and the input data of the latter is implicit feedback data, such as clicking, purchasing, watching and other behaviors of the user. Collaborative filtering recommendation algorithms face serious data sparseness and cold start problems because it is difficult to obtain explicit feedback data for users in some cases. Unlike explicit feedback data, implicit feedback data is widely available and resource-rich. Although the advantages of implicit feedback data are quite clear, there is a lack of negative feedback data in this type of data. Based on this, researchers have proposed a typical bayesian personalized ranking algorithm to solve this problem.
At present, because items, scoring habits, browsing time lengths and the like of interest of registered users in the virtual browsing system have uncertainty, the user-item scoring matrix in the system database is changed frequently, and the sparseness of part of user scoring data sets is reduced along with the increase of old users in the system. At present, researchers have conducted extended research on Bayesian personalized ranking algorithms and proposed various recommendation methods. However, the above-described extension method focuses only on the unscored items of the user, and does not consider the negative feedback data contained in the historical scored items of the user and the explicit characteristic information in the historical scored data, and it is apparent that the existing method causes unbalance of the positive feedback data and the negative feedback data and waste of data resources.
Disclosure of Invention
In order to solve the problem of poor performance of a recommendation system caused by unbalance of positive feedback data and negative feedback data and waste of data resources in the existing recommendation method, the invention provides a post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics.
The technical scheme adopted by the invention is as follows:
the post-fusion personalized recommendation model based on the explicit feedback characteristics and the implicit feedback characteristics comprises an explicit characteristic extraction module, an implicit characteristic extraction module and an integral characteristic extraction module, wherein the explicit characteristic extraction module and the implicit characteristic extraction module are respectively connected with the integral characteristic extraction module;
the explicit feature extraction module is used for receiving data of the user feature matrix and the project feature matrix, extracting explicit feedback features of the user and the project, and reconstructing a predicted scoring matrix;
the implicit characteristic extraction module is used for receiving the explicit scoring data and the implicit feedback data of the user, extracting the implicit feedback characteristics of the user and the project and reconstructing a predicted sequencing scoring matrix;
and the integral feature extraction module is used for carrying out weighted summation on the predicted scoring matrix and the predicted sorting scoring matrix to obtain a final predicted sorting scoring matrix.
Further, the implicit feature extraction module adopts a BPR model;
the BPR model is used for collecting D from partial sequences s Extracting characteristics of users and projects;
the partial order set D s Is defined as follows:
wherein item i represents any item marked by user u, item j represents any item not marked by user u, IV (u) represents a set of items marked by user u once,representing a set of items that user u has not marked, the triplet (u, i, j) represents that user u prefers item i over item j.
Further, the implicit feature extraction module adopts an IBPR model;
the IBPR model is used for expanding the data set of the pair data input by the BPR model, increasing the definition of the user on the historical scoring item pair and expanding the partial order set D of the part R The definition is as follows:
D R ＝{(u，i，j)li∈Ⅳ(u)and j∈Ⅳ(u)and r ui ＞r uj }
wherein item i and item j represent two of the items marked by user u, N (u) represents the set of items that user u has marked, r ui and ruj Representing the historical scoring values of user u on item i and item j, respectively, the triplet (u, i, j) represents that user u prefers item i over item j, and therefore the input partial order set D is represented as:
D＝D S ∪D R
and part of user history scoring item pairs are added in the partial order set D.
Further, the objective function of the IBPR model is:
wherein sigma (x) is a Sigmoid function, lambda is a regularization parameter, theta is a parameter set of the IBPR model, and theta= { b u ，b i ，p u ，q i ，q j }，For obtaining the relationship between user u and two items i and j,/and j>Is defined as follows:
wherein , and />Respectively representing the sorting priority of the user u predicted by the IBPR model to the item i and the item j, < >>b i Bias term, r, representing item i max and rmin Representing the maximum and minimum, respectively, of the user scores in the dataset, p u Feature vector, q representing user u i Feature vector representing item i, < >>Represents q i Is a transpose of (a).
Further, the explicit feature extraction module adopts a BiasSVD model;
the scoring prediction formula of any user u of the BiasSVD model on the item i is adoptedWherein μ is the training setScore value of historical score data, b i and bu Representing the bias term of item i and the bias term of user u, p, respectively u Feature vector, q representing user u i Feature vector representing item i, < >>Represents q i Is a transpose of (a).
Further, the integral feature extraction module adopts an IBPR SVD model;
the IBPR_SVD model is fused with the BiasSVD model and the IBPR model;
for user u, the ranking score of any unlabeled item i is predictedThe following formula is used for calculation:
wherein ,representing the score value of user u on item i predicted by BiasSVD model,/L>Representing the sorting score of the user u on the item i, which is predicted by the IBPR model, wherein alpha is a compromise parameter, and the value range of alpha is [0,1]]。
Further, the optimal value of the compromise parameter alpha is 0.9.
A post-fusion personalized recommendation method based on explicit and implicit feedback features comprises the following steps:
step one: acquiring explicit scoring data and implicit feedback data of a user from a user feedback database;
step two: extracting explicit feedback characteristics of a user and a project by using a BiasSVD model, learning a user characteristic matrix and a project characteristic matrix according to historical scoring data of the user, and reconstructing a predicted scoring matrix through the learned characteristic matrix;
step three: extracting implicit feedback characteristics of a user and a project by using an IBPR model, constructing a diagonal project preference set of the IBPR model by using explicit scoring data and implicit feedback data of the user, obtaining another group of user feature matrixes and project feature matrixes by training the IBPR model on the extended diagonal project set, predicting the ranking scores of all unlabeled projects in the dataset according to a ranking score prediction formula, and further reconstructing a ranking score matrix of corresponding user-project prediction;
step four: weighting and summing a predicted scoring matrix obtained by the BiasSVD model and a predicted sorting scoring matrix obtained by the IBPR model to obtain a final predicted sorting scoring matrix, wherein the higher the sorting score is, the higher the probability of obtaining recommendation is;
step five: and sequencing all sequencing scores obtained by the user prediction from high to low to obtain a recommendation list, and recommending the top IV items sequenced in the front to the user.
The invention has the beneficial effects that:
1. based on the assumption that the user prefers the items with higher scoring values, on the basis of a Bayesian personalized ranking model, user scoring item pairs are introduced, a new objective function is defined on a set, an improved Bayesian personalized ranking model IBPR is provided, the problem of lack of negative feedback in implicit feedback data is relieved by the IBPR model, and the anti-noise performance of the model is improved;
2. the invention provides a post-fusion personalized recommendation model and method based on explicit and implicit feedback features, which combines an IBPR model and a BiasSVD model, utilizes the BiasSVD model to extract explicit feedback features of users and projects, utilizes the IBPR model to extract implicit feedback features of the users and the projects, reconstructs a user-project ranking scoring matrix according to the extracted implicit feedback features, reconstructs the user-project scoring matrix according to the explicit feedback features, fuses the reconstructed matrices, and the fused matrices are the final ranking scoring matrix.
Drawings
FIG. 1 is a frame diagram of a post-fusion personalized recommendation model based on explicit and implicit feedback features provided in embodiment 1 of the present invention;
FIG. 2 is a graphical comparison of a BPR model and a proposed IBPR model generation pair-wise partial order set;
FIG. 3 is a graph showing the variation of MAP/MRR index with the value of the compromise parameter alpha on the public dataset Movielens 100K by the IBPR_SVD model;
FIG. 4 is a graph showing the variation of MAP/MRR index along with the value of the compromise parameter alpha on the public data set FilmTrust by the IBPR_SVD model;
FIG. 5 is a schematic diagram of an attention mechanism module;
FIG. 6 is a graph showing the MAP index of the IBPR_SVD model as a function of the number of scores;
FIG. 7 is a graph showing the variation of MRR index of the IBPR_SVD model with the number of scores;
fig. 8 is a flowchart of a post-fusion personalized recommendation method based on explicit and implicit feedback features provided in embodiment 2 of the present invention.
Detailed Description
The invention will be described in detail below with reference to the drawings and the detailed description.
Example 1:
the embodiment provides a post-fusion personalized recommendation model based on explicit and implicit feedback features, which comprises an explicit feature extraction module, an implicit feature extraction module and an integral feature extraction module, wherein the explicit feature extraction module and the implicit feature extraction module are respectively connected with the integral feature extraction module. The explicit feature extraction module is used for receiving data of the user feature matrix and the project feature matrix, extracting explicit feedback features of the user and the project, and reconstructing a predicted scoring matrix; the implicit characteristic extraction module is used for receiving the explicit scoring data and the implicit feedback data of the user, extracting the implicit feedback characteristics of the user and the project and reconstructing a predicted sequencing scoring matrix; the overall feature extraction module is used for carrying out weighted summation on the predicted scoring matrix and the predicted sorting scoring matrix to obtain a final predicted sorting scoring matrix.
Referring to fig. 1, in fig. 1, a part identifying the biasvd algorithm is an explicit feature extraction module, a part identifying the IBPR algorithm is an implicit feature extraction module, and a right part of fig. 1 is an overall feature extraction module. In fig. 1, U is a user feature matrix in the biasvd algorithm, V is an item feature matrix in the biasvd algorithm, and the predicted scoring matrix is U 1 For the user characteristic matrix in the IBPR algorithm, V 1 For the feature matrix of the item in the IBPR algorithm, the predicted sorting scoring matrix is +.>
In this embodiment, the explicit feature extraction module adopts a biasvd model (Bias Singular ValueDecomposition), the implicit feature extraction module adopts a BPR model, and may also adopt an IBPR model, and the overall feature extraction module adopts an ibpr_svd model.
Based on the assumption that the browsing user prefers the items with high scoring values, the embodiment increases the definition of the model on the explicit scoring data on the basis of the Bayesian personalized ranking model, and provides an improved Bayesian personalized ranking IBPR (Improved Bayesian Personalized Ranking, IBPR) model for extracting the implicit feedback characteristics of the browsing user and the items. The model fully utilizes the existing browsing user feedback data and introduces the definition of the negative feedback data, thereby relieving the problem of lack of negative feedback in the implicit feedback data and improving the noise resistance of the model.
In order to alleviate the cold start problem, the embodiment adopts a BiasSVD model to extract the explicit feedback characteristics of users and projects, reconstructs a user-project ordering scoring matrix according to the extracted implicit feedback characteristics, reconstructs a user-project scoring matrix according to the explicit feedback characteristics, fuses the reconstructed matrices, and provides a post-fusion personalized recommendation model based on the explicit and implicit feedback characteristics. Since the model combines the IBPR model and the biasvd model, the model is simply referred to as ibpr_svd (Improved Bayesian PersonalizedRanking Singular Value Decomposition, ibpr_svd).
From the definition of the BPR (Bayesian Personalized Ranking, BPR) model, the model is derived from the partial order set D only S Features of the user and the item are extracted, and pairs of items marked by the user are ignored. D (D) S Is defined as follows:
wherein item i represents any item marked by user u, and item j represents any item not marked by user u. IV (u) represents the set of items that user u has marked,representing a collection of items that user u has not marked. The triplet (u, i, j) indicates that user u prefers item i over item j.
Since in browsing systems, it is common for one user to evaluate several or even more browsing items simultaneously, it can be intuitively seen from the value of the score: the user prefers items with higher scoring values. Based on this, the present embodiment proposes an IBPR model based on the assumption that the user prefers items with high score values. Different from the BPR model, the IBPR model provided by the embodiment expands the pair-level data set input by the BPR model, increases the definition of the user on the history scoring item pair, and expands the partial order set D of the part R The definition is as follows:
D R ＝{(u，i，j)|i∈Ⅳ(u)and j∈Ⅳ(u)and r ui ＞r uj } (2)
wherein item i and item j represent two of the items marked by user u, N (u) represents the set of items that user u has marked, r ui and ruj The historical scoring values for user u on item i and item j, respectively. The triplet (u, i, j) indicates that user u prefers item i over item j. Thus, the partial order set D of IBPR model inputs can be expressed as:
D＝D S ∪D R (3)
part of user history scoring item pairs are added in the partial order set D.
The BPR model versus data set and IBPR model versus data set of this embodiment refer to fig. 2.
FIG. 2 illustrates the conversion of user feedback data into a pair-wise partial order set D by a BPR model and an IBPR model, respectively S And D. The leftmost user-item scoring matrix in FIG. 2 is a set of scoring records for a user on an item, with the score reflecting the user's preference, and the higher the score, the more liked the user. The question mark indicates that the user never scored the item. The BPR model and IBPR model generate a preferred item pair u between two different items according to the user's tagging situation: i > u The plus sign in the partial order set indicates that user u prefers item i over item j, while the minus sign is the exact opposite, which indicates that user prefers item j over item i, and the question mark indicates that the user's relative preference for both items cannot be judged.
As can be seen from FIG. 2, the partial user history scoring item pairs are added in the partial order set of the IBPR model relative to the BPR model, and the triplet < u 1 ，i 2 ，i 3 For example, in the BPR model, user u cannot be determined 1 For item i 2 And item i 3 Is a preferred relationship of (a). Due to user u 1 In item i 2 And item i 3 Scoring records on 5 and 3, respectively, the IBPR model then considers relative to item i 3 User u 1 Preference i 2 . If the user's score values are the same on both items,the IBPR model cannot determine the preference relationship between two items.
Based on the above analysis, the objective function of the IBPR model is:
wherein sigma (x) is a Sigmoid function, lambda is a regularization parameter, theta is a parameter set of the model, and theta= < b u ，b i ，p u ，q i ，q j }。For obtaining the relationship between user u and two items i and j,/and j>Is defined as follows:
wherein , and />Respectively representing the order priorities of the user u predicted by the model to the item i and the item j,b i bias term, r, representing item i max and rmin Representing the maximum and minimum, respectively, of the user scores in the dataset, p u Feature vector, q representing user u i Feature vector representing item i, < >>Represents q i Is a transpose of (a). IBPR model is optimized and solved by random gradient descent methodAnd (5) solving.
In this embodiment, the explicit feature extraction module adopts a biasvd model, that is, adopts a biasvd model to extract explicit feedback features of the user and the item, and the scoring prediction formula of any user u of the biasvd model on the item i adoptsWherein mu is the scoring value of the historical scoring data of the training set, b i and bu Representing the bias term of item i and the bias term of user u, p, respectively u Feature vector, q representing user u i Feature vector representing item i, < >>Represents q i Is a transpose of (a).
The integral feature extraction module of the embodiment adopts an IBPR_SVD model, and the IBPR_SVD model fuses the BiasSVD model and the IBPR model.
For user u, the predicted ranking score of any one of the unlabeled items iThe calculation can be made by the following formula:
wherein ,representing the score value of user u on item i predicted by BiasSVD model,/L>And (5) representing the ranking scores of the users u on the item i, which are predicted by the IBPR model. Alpha is a compromise parameter, and the value range of alpha is [0,1]。
Because the quality of the value of the compromise parameter α directly affects the recommendation accuracy of the algorithm proposed in this embodiment, before verifying the performance of the ibpr_svd algorithm, it is first necessary to determine the optimal value of the compromise parameter α. The larger the compromise parameter alpha is, the larger the influence of the IBPR model on the IBPR_SVD algorithm is, and the smaller the influence of the BiasSVD model is; the smaller the compromise parameter alpha is, the smaller the contribution of the IBPR model to the IBPR_SVD algorithm is, and the contribution of the BiasSVD model is increased. When α=0, the ibpr_svd algorithm degenerates to the biasvd model, and when α=1, the ibpr_svd algorithm degenerates to the IBPR model. The specific value of the compromise parameter alpha is determined through experiments.
In this embodiment, the most suitable value of the compromise parameter α is determined through experiments, the experiments are performed on two public data sets, movieens 100K and FilmTrust, the compromise parameter α is a variable, the value is [0,1], the step size is 0.1, map and MRR are used as evaluation indexes of the experiments, the iteration number is set to 1000, and the feature vector dimensions of the ibpr model and the biasvd model are both set to 10. The performance curves of the IBPR SVD algorithm as a function of the compromise parameter a on the two data sets movieens 100K and FilmTrust are shown in fig. 3 and 4. As can be seen from fig. 3 and fig. 4, the optimal value of the compromise parameter is 0.9.
Unlike the existing method in which all the history scoring items are used as positive feedback data, the embodiment defines negative feedback data in the history scoring items; and (3) connecting a collaborative filtering algorithm BiasSVD based on historical scoring data and a collaborative filtering algorithm IBPR based on ranking scoring by adopting a compromise parameter, and jointly extracting the explicit characteristics and the implicit characteristics of the user and the project. Historical scoring data and implicit feedback data in the data set are fully utilized, and the cold start problem of the recommendation system is relieved.
Example 2:
the embodiment provides a post-fusion personalized recommendation method based on explicit and implicit feedback characteristics, and reference is made to fig. 8. The post-fusion personalized recommendation method based on explicit and implicit feedback features fuses a BiasSVD model and an IBPR model, and comprises the following steps:
step one: acquiring explicit scoring data and implicit feedback data of a user from a user feedback database;
step two: extracting explicit feedback characteristics of a user and a project by using a BiasSVD model, learning a user characteristic matrix and a project characteristic matrix according to historical scoring data of the user, and reconstructing a predicted scoring matrix through the learned characteristic matrix;
step three: extracting implicit feedback characteristics of a user and a project by using an IBPR model, constructing a diagonal project preference set of the IBPR model by using explicit scoring data and implicit feedback data of the user, obtaining another group of user feature matrixes and project feature matrixes by training the IBPR model on the extended diagonal project set, predicting the ranking scores of all unlabeled projects in the dataset according to a ranking score prediction formula, and further reconstructing a ranking score matrix of corresponding user-project prediction;
step four: weighting and summing a predicted scoring matrix obtained by the BiasSVD model and a predicted sorting scoring matrix obtained by the IBPR model to obtain a final predicted sorting scoring matrix, wherein the higher the sorting score is, the higher the probability of obtaining recommendation is;
step five: and sequencing all sequencing scores obtained by the user prediction from high to low to obtain a recommendation list, and recommending the top IV items sequenced in the front to the user.
The post-fusion personalized recommendation method based on the explicit feedback and implicit feedback features combines an IBPR model and a BiasSVD model. In the embodiment, the BiasSVD model is used as an explicit feature extraction module to extract explicit feedback features of users and projects, the predicted scoring matrix is reconstructed, the IBPR model is used as an implicit feature extraction module to extract implicit feedback features of the users and the projects, the predicted ranking scoring matrix is reconstructed, and the IBPR SVD model is used as an integral feature extraction module to fuse the predicted scoring matrix with the predicted ranking scoring matrix, so that the final predicted ranking scoring matrix is obtained. For specific methods, reference may be made to the processing procedure of embodiment 1 based on post-fusion personalized recommendation models of explicit feedback and implicit feedback features. The post-fusion personalized recommendation method based on the explicit feedback and implicit feedback features fully utilizes historical scoring data and implicit feedback data in the data set, and alleviates the problem of cold start of a recommendation system.
Workflow of the present embodiment, input: the feature vector dimension f of the IBPR_SVD model, the learning rate gamma, the regularization parameter lambda, the iteration times interfaces, the rank-partial order set D, the number of users m in the data set and the number of items n; output predicted ranking scoring matrixInitializing preRmse and parameter sets in a BiasSVD model and a BPR model; the operation is started until the iteration stops.
To explore the most efficient way to fuse biasvd and IBPR, this example designed 13 different fusion methods to compare. In order to more clearly describe the above method, the present embodiment assumes that the user and project feature vectors of the biasvd model are U and V, respectively, and the user and project feature vectors of the IBPR model are P and Q, respectively. The 13 fusion modes are defined as follows:
1) Ibpr_svd_am. This embodiment designs an attention mechanism module to assign weights to biasvd and IBPR as shown in fig. 5. In the IBPR SVD AM method, the corresponding attention weights may be different for different users. In FIG. 5, inputs a (u) and b (u) represent the influence of BiasSVD and IBPR on user u, respectively, and the vectors at the outputRepresenting the user attention weight vector due to the influence of biasvd, +.>Representing the user attention weight vectors due to the influence of IBPR, respectively. The present embodiment designs a loss function to calculate the vector +.> and />
wherein ,represent training set, r ui Representing historical scoring data, λ is a regularization parameter, and the present embodiment uses a gradient descent method to optimize the above-described loss function.
2) Ibpr_svd_max. The method takes the maximum value in the BiasSVD and IBPR prediction results as the final prediction result.
3) Ibpr_svd_min. The method takes the minimum value in the BiasSVD and IBPR prediction results as the final prediction result.
4) Ibpr_svd_mul. The method takes the product of BiasSVD and IBPR prediction results as the final prediction result.
5) Ibpr_svd_ (u+p) (v+q). To try more possibilities, the method reassembles the biasvd and IBPR model feature sets. The method comprises the steps of firstly calculating the sum of preference characteristics of two model users and the sum of attribute characteristics of items, and then carrying out scoring prediction on unscored items.
6) Ibpr_svd (uq+vp). The method recombines the features of the two models, exchanges the BiasSVD and IBPR model item features, and calculates corresponding dot product.
7) Ibpr_svd_uq. The method extracts the user characteristics of BiasSVD and the project characteristics of the IBPR model to recommend.
8) Ibpr_svd_vp. The method extracts project features of BiasSVD and user features of the IBPR model for recommendation.
9) Ibpr_svd (u+p) V. The method adopts the user characteristic sum of the two models and the project characteristic of the BiasSVD model to recommend.
10 Ibpr_svd (u+p) Q. The method adopts the user characteristic sum of the two models and the project characteristic of the IBPR model to recommend.
11 Ibpr_svd_u (v+q). The method adopts the user characteristics of the BiasSVD model and the project characteristics of the two models to recommend.
12 Ibpr_svd_p (v+q). The method adopts the user characteristics of the IBPR model and the project characteristics of the two models to recommend.
13 Ibpr_svd. The method is the method proposed by the embodiment, the BiasSVD and the IBPR model are fused by adopting a compromise parameter alpha, and the compromise parameter is set to be 0.9.
According to the above description, the performance of 13 fusion methods was tested on two public data sets, movieens 100K and film trust, and the evaluation index was selected from the group consisting of precision@3, recall@3, precision@5, recall@5, MAP and MRR, and the experimental results are shown in table 1. From the experimental results, the fusion method proposed in this embodiment is significantly superior to other fusion methods.
TABLE 1
To illustrate the effectiveness of the method IBPR_SVD presented in this example, a number of experiments were performed on 5 published data sets, movietens 100K, movielens 1M, filmTrust, ciao and Hetrec-movietens-2 k, respectively. The effectiveness of the IBPR_SVD method is further verified by comparing with 4 novel methods. The indexes adopted in the experiment are precision@3, recall@3, precision@5, recall@5, MAP and MRR, respectively, and the experimental results are shown in Table 2. Experimental results show that the performance of the IBPR_SVD is obviously superior to that of other methods, and the method provided by the embodiment is proved to be effective again.
TABLE 2
/>
/>
The models to which this example is compared are published in the top journal of the information recommendation area, including knowledges-Based Systems (KBS) and Information Sciences (Inf. Sci). Comparative model: RBPR (Rating Bayesian personalized ranking), SPR (Similarity pairwise ranking), BPRN (Bayesian personalized ranking algorithm based on multiple-layer neighborhoods) and MSBPR (multi-pairwise preference and similarity based BPR).
The present embodiment explores the performance of the ibpr_svd method in the case of a user cold start, and the new user cold start dataset used in the present embodiment is artificially generated from the FilmTrust dataset. The specific method comprises the following steps: the cold start data set is generated by changing the scoring number of the users, the scoring number of the users of the data set takes the value range of [3,19], and the step length is 2. The number of scores of users is different from one dataset to another. The comparative method of this example selects the SPR method, and the evaluation index selects the MAP and the MRR. As shown in fig. 6 and fig. 7, it can be seen from fig. 6 and fig. 7 that the ibpr_svd method proposed in the present embodiment can alleviate the problem of cold start to some extent.
The foregoing description is only of the preferred embodiments of the present invention, and is not intended to limit the invention, but rather should be construed in scope without departing from the technical scope of the invention.

Claims (8)

1. The post-fusion personalized recommendation model based on the explicit and implicit feedback features is characterized by comprising an explicit feature extraction module, an implicit feature extraction module and an integral feature extraction module, wherein the explicit feature extraction module and the implicit feature extraction module are respectively connected with the integral feature extraction module;
the explicit feature extraction module is used for receiving data of the user feature matrix and the project feature matrix, extracting explicit feedback features of the user and the project, and reconstructing a predicted scoring matrix;
the implicit characteristic extraction module is used for receiving the explicit scoring data and the implicit feedback data of the user, extracting the implicit feedback characteristics of the user and the project and reconstructing a predicted sequencing scoring matrix;
and the integral feature extraction module is used for carrying out weighted summation on the predicted scoring matrix and the predicted sorting scoring matrix to obtain a final predicted sorting scoring matrix.
2. The post-fusion personalized recommendation model based on explicit and implicit feedback features of claim 1, wherein the implicit feature extraction module employs a BPR model;
the BPR model is used for collecting D from partial sequences s Extracting characteristics of users and projects;
the partial order set D s Is defined as follows:
wherein item i represents any item marked by user u, item j represents any item not marked by user u, N (u) represents a set of items marked by user u once,representing a set of items that user u has not marked, the triplet (u, i, j) represents that user u prefers item i over item j.
3. The post-fusion personalized recommendation model based on explicit and implicit feedback features of claim 2, wherein the implicit feature extraction module employs an IBPR model;
the IBPR model is used for expanding the data set of the pair data input by the BPR model, increasing the definition of the user on the historical scoring item pair and expanding the partial order set D of the part R The definition is as follows:
D R ＝{(u，i，j)|i∈N(u)and j∈N(u)and r ui ＞r uj }
wherein item i and item j represent two of the items marked by user u, N (u) represents the set of items that user u has marked, r ui and ruj Representing the historical scoring values of user u on item i and item j, respectively, the triplet (u, i, j) represents that user u prefers item i over item j, and therefore the input partial order set D is represented as:
D＝D S ∪D R
and part of user history scoring item pairs are added in the partial order set D.
4. The post-fusion personalized recommendation model based on explicit and implicit feedback features of claim 3, wherein the objective function of the IBPR model is:
wherein sigma (x) is a Sigmoid function, lambda is a regularization parameter, theta is a parameter set of the IBPR model, and theta= { b u ，b i ，p u ，q i ，q j }，For obtaining the relationship between user u and two items i and j,/and j>Is defined as follows:
wherein , and />Respectively representing the order priority of the user u predicted by the IBPR model to the item i and the item j,b i bias term, r, representing item i max and rmin Representing the maximum and minimum, respectively, of the user scores in the dataset, p u Feature vector, q representing user u i Feature vector representing item i, < >>Represents q i Is a transpose of (a).
5. The post-fusion personalized recommendation model based on explicit and implicit feedback features of claim 4, wherein the explicit feature extraction module employs a biasvd model;
the scoring prediction formula of any user u of the BiasSVD model on the item i is adoptedWherein mu is the scoring value of the historical scoring data of the training set, b i and bu Representing the bias term of item i and the bias term of user u, p, respectively u Feature vector, q representing user u i Feature vector representing item i, < >>Represents q i Is a transpose of (a).
6. The post-fusion personalized recommendation model based on explicit and implicit feedback features of claim 5, wherein the global feature extraction module employs an IBPR SVD model;
the IBPR_SVD model is fused with the BiasSVD model and the IBPR model;
for user u, the ranking score of any unlabeled item i is predictedThe following formula is used for calculation:
wherein ,representing the score value of user u on item i predicted by BiasSVD model,/L>Representing the sorting score of the user u on the item i, which is predicted by the IBPR model, wherein alpha is a compromise parameter, and the value range of alpha is [0,1]]。
7. The post-fusion personalized recommendation model based on explicit and implicit feedback features of claim 6, wherein the optimal value of the compromise parameter α is 0.9.
8. The post-fusion personalized recommendation method based on the explicit and implicit feedback characteristics is characterized by comprising the following steps of:
step one: acquiring explicit scoring data and implicit feedback data of a user from a user feedback database;
step two: extracting explicit feedback characteristics of a user and a project by using a BiasSVD model, learning a user characteristic matrix and a project characteristic matrix according to historical scoring data of the user, and reconstructing a predicted scoring matrix through the learned characteristic matrix;
step three: extracting implicit feedback characteristics of a user and a project by using an IBPR model, constructing a diagonal project preference set of the IBPR model by using explicit scoring data and implicit feedback data of the user, obtaining another group of user feature matrixes and project feature matrixes by training the IBPR model on the extended diagonal project set, predicting the ranking scores of all unlabeled projects in the dataset according to a ranking score prediction formula, and further reconstructing a ranking score matrix of corresponding user-project prediction;
step four: weighting and summing a predicted scoring matrix obtained by the BiasSVD model and a predicted sorting scoring matrix obtained by the IBPR model to obtain a final predicted sorting scoring matrix, wherein the higher the sorting score is, the higher the probability of obtaining recommendation is;
step five: and sequencing all sequencing scores predicted by the user from high to low to obtain a recommendation list, and recommending the top N items sequenced in the front to the user.
CN202310619062.XA 2023-05-29 2023-05-29 Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics Pending CN116701861A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310619062.XA CN116701861A (en) 2023-05-29 2023-05-29 Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310619062.XA CN116701861A (en) 2023-05-29 2023-05-29 Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics

Publications (1)

Publication Number Publication Date
CN116701861A true CN116701861A (en) 2023-09-05

Family

ID=87823118

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310619062.XA Pending CN116701861A (en) 2023-05-29 2023-05-29 Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics

Country Status (1)

Country Link
CN (1) CN116701861A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117974330A (en) * 2024-03-28 2024-05-03 华侨大学 Internet insurance score prediction method and device based on hybrid model

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117974330A (en) * 2024-03-28 2024-05-03 华侨大学 Internet insurance score prediction method and device based on hybrid model

Similar Documents

Publication Publication Date Title
CN111291261A (en) Cross-domain recommendation method integrating label and attention mechanism and implementation system thereof
CN107220365A (en) Accurate commending system and method based on collaborative filtering and correlation rule parallel processing
CN108280114A (en) A kind of user&#39;s literature reading interest analysis method based on deep learning
CN110930219B (en) Personalized merchant recommendation method based on multi-feature fusion
CN111858972B (en) Movie recommendation method based on family knowledge graph
CN116701861A (en) Post-fusion personalized recommendation model and method based on explicit and implicit feedback characteristics
CN112612820A (en) Data processing method and device, computer readable storage medium and processor
CN113742464A (en) News event discovery algorithm and device based on heterogeneous information network
CN111259264B (en) Time sequence scoring prediction method based on generation countermeasure network
CN113343077A (en) Personalized recommendation method and system integrating user interest time sequence fluctuation
CN114154565A (en) Click rate prediction method and device based on multi-level feature interaction
CN114781503A (en) Click rate estimation method based on depth feature fusion
CN113221003B (en) Mixed filtering recommendation method and system based on dual theory
Zhu et al. Multimodal sparse linear integration for content-based item recommendation
CN114564594A (en) Knowledge graph user preference entity recall method based on double-tower model
KR20220125422A (en) Method and device of celebrity identification based on image classification
CN112817442A (en) Situation information classification recommendation system and method under multi-task condition based on FFM
Feng et al. Personalized recommendation with hybrid feedback by refining implicit data
CN115730139A (en) Short video recommendation method and device based on heterogeneous multi-interest modeling
CN114238439A (en) Task-driven relational data view recommendation method based on joint embedding
CN112203152A (en) Multi-modal confrontation learning type video recommendation method and system
Xue et al. An explicit feedback recommendation algorithm based on subjective and objective evaluation transformation model
Wang et al. BERT-based aggregative group representation for group recommendation
Nie et al. Movie recommendation using unrated data
AU2021101719A4 (en) Logo-based document image retrieval and method thereof

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