Post-fusion personalized recommendation model and method based on explicit and implicit feedback features

Abstract

The application particularly relates to a post-fusion personalized recommendation model and method based on explicit and implicit feedback features, the model comprising an explicit feature extraction module, an implicit feature extraction module and an overall feature extraction module, the explicit feature extraction module and the implicit feature extraction module being connected with the overall feature extraction module respectively; the method fuses an IBPR model and a BiasSVD model, performs weighted summation on a predicted score matrix obtained by the BiasSVD model and a predicted ranking score matrix obtained by the IBPR model, obtains a final predicted ranking score matrix, ranks all predicted ranking scores of users from high to low, and recommends the first items with higher ranking to the users. The application extracts explicit feedback features by using the BiasSVD model, extracts implicit feedback features by using the IBPR model, and fully utilizes historical score data and implicit feedback data in a data set, thereby relieving a cold start problem of a recommendation system and improving the performance of the recommendation system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of recommendation system, information retrieval and data mining, and relates to a post-fusion personalized recommendation model and method based on explicit and implicit feedback features. BACKGROUND

[0002] With the explosive growth of data, recommendation systems are used to solve the problem of information overload. In a recommendation system, the recommendation algorithm is the most core and key part, which largely determines the pros and cons of the performance of the recommendation system. At present, there are many different recommendation algorithms, among which the most widely used recommendation algorithm is the collaborative filtering recommendation algorithm.

[0003] According to different input data, the collaborative filtering recommendation algorithm can be divided into two categories: a collaborative filtering recommendation algorithm based on historical ratings and a collaborative filtering recommendation algorithm based on ranking ratings. The input data of the former is explicit feedback data, such as user historical rating data, and the input data of the latter is implicit feedback data, such as user clicks, purchases, views and the like. Since it is difficult to obtain explicit feedback data of users in some cases, the collaborative filtering recommendation algorithm faces serious data sparsity problems and cold start problems. Unlike explicit feedback data, implicit feedback data is widely sourced and abundant. Although the advantages of implicit feedback data are very prominent, 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.

[0004] At present, due to the uncertainty of the registered users' interest in the project, rating habits and browsing time in the virtual browsing system, the user-item rating matrix in the system database changes frequently, and as the number of old users in the system increases, the sparsity of the user rating data set will also decrease. At present, researchers have carried out extended research on the Bayesian personalized ranking algorithm and have proposed a variety of recommendation methods. However, the above-mentioned extension methods only focus on the un-rated items of the user, and do not consider the negative feedback data contained in the user's historical rating items and the explicit feature information in the historical rating data. Obviously, the existing method causes imbalance between positive feedback data and negative feedback data and waste of data resources. SUMMARY

[0005] In order to solve the problem of poor performance of the recommendation system caused by the imbalance between positive feedback data and negative feedback data and the waste of data resources in the existing recommendation method, the application provides a post-fusion personalized recommendation model and method based on explicit and implicit feedback features.

[0006] The technical scheme adopted by the application is:

[0007] The post-fusion personalized recommendation model based on explicit and implicit feedback features comprises an explicit feature extraction module, an implicit feature extraction module and an overall feature extraction module, the explicit feature extraction module and the implicit feature extraction module are connected with the overall feature extraction module respectively;

[0008] The explicit feature extraction module is used for receiving data of a user feature matrix and an item feature matrix, extracting explicit feedback features of the user and the item, and reconstructing a predicted rating matrix.

[0009] The implicit feature extraction module is used for receiving explicit rating data and implicit feedback data of the user, extracting implicit feedback features of the user and the item, and reconstructing a predicted ranking rating matrix.

[0010] The overall feature extraction module is used for weighted sum of the predicted rating matrix and the predicted ranking rating matrix to obtain a final predicted ranking rating matrix.

[0011] Further, the implicit feature extraction module adopts a BPR model.

[0012] The BPR model is used for extracting features of the user and the item from a partial order set D s .

[0013] The partial order set D s is defined as follows:

[0014]

[0015] Wherein, the item i represents any one item marked by the user u, the item j represents any one item not marked by the user u, IV(u) represents a set of items marked by the user u, IV(u) represents a set of items not marked by the user u, and the triple (u, i, j) represents that the user u prefers the item i to the item j.

[0016] Further, the implicit feature extraction module adopts an IBPR model.

[0017] The IBPR model is used for expanding the pair-level data set input by the BPR model, increasing the definition of the user on the historical rating item pair, and the expanded partial order set D R is defined as follows:

[0018] D R ={(u, i, j) | i ∈ IV(u) and j ∈ IV(u) and r ui > r uj}

[0019] where, item i and item j represent two items that user u has marked, N(u) represents a set of items that user u has marked, r ui and r uj respectively represent the historical rating value of user u on item i and item j, and the triplet (u, i, j) represents that user u prefers item i to item j, and thus the input partial order set D is represented as:

[0020] D = D S ∪ D R

[0021] The partial order set D is added with part of the user historical rating item pairs.

[0022] Further, the objective function of the IBPR model is:

[0023]

[0024] where, σ(x) is a Sigmoid function, λ is a regularization parameter, Θ is a parameter set of the IBPR model, Θ = {b u , b i , p u , q i , q j}, is used to obtain the relationship between user u and two items i and j, is defined as follows:

[0025]

[0026] where, and respectively represent the ranking priority of user u on item i and item j predicted by the IBPR model, b i represents a bias term of item i, r max and r min respectively represent the maximum value and the minimum value of the user rating in the data set, p u represents a feature vector of user u, q i represents a feature vector of item i, represents the transpose of q i .

[0027] Further, the explicit feature extraction module adopts a BiasSVD model;

[0028] The rating prediction formula of the BiasSVD model for any user u on item i is where, μ is a rating value of the historical rating data of the training set, b i and b urespectively represent the bias term of item i and the bias term of user u, p u represent the feature vector of user u, q i represent the feature vector of item i, represent q i transpose of q.

[0029] Further, the overall feature extraction module adopts an IBPR SVD model;

[0030] The IBPR SVD model combines the BiasSVD model and the IBPR model;

[0031] For user u, the ranking score of any unlabeled item i predicted is calculated by the following formula:

[0032]

[0033] wherein, represents the rating value of user u on item i predicted by the BiasSVD model, represents the ranking score of user u on item i predicted by the IBPR model, and a is a compromise parameter, and the value range of a is [0, 1].

[0034] Further, the optimal value of the compromise parameter a is 0.9.

[0035] A post-fusion personalized recommendation method based on explicit and implicit feedback features, comprising the following steps:

[0036] Step one: obtaining explicit rating data and implicit feedback data of users from a user feedback database;

[0037] Step two: extracting explicit feedback features of users and items by using the BiasSVD model, learning user feature matrices and item feature matrices according to historical rating data of users, and reconstructing a predicted rating matrix through the learned feature matrices;

[0038] Step three: extracting implicit feedback features of users and items by using the IBPR model, constructing a pair-level item preference set of the IBPR model through explicit rating data and implicit feedback data of users, obtaining another group of user feature matrices and item feature matrices by training the IBPR model on the extended pair-level item set, predicting the ranking scores of all unlabeled items in the data set according to a ranking score prediction formula, and further reconstructing a corresponding user-item predicted ranking score matrix;

[0039] Step four: the predicted rating matrix obtained by the BiasSVD model and the predicted ranking score matrix obtained by the IBPR model are weighted and summed to obtain a final predicted ranking score matrix, and the higher the ranking score is, the greater the probability of obtaining recommendation is;

[0040] Step five: all the ranking scores predicted by the user are sorted from high to low to obtain a recommendation list, and the top four items are recommended to the user.

[0041] The beneficial effects of the present application are:

[0042] 1. The present application is based on the assumption that users prefer items with higher rating values, and introduces user rating item pairs on the basis of the Bayesian personalized ranking model, defines a new objective function on the set, and proposes an improved Bayesian personalized ranking model IBPR, which alleviates the problem of lack of negative feedback in implicit feedback data and improves the noise resistance of the model;

[0043] 2. The present application proposes a post-fusion personalized recommendation model and method based on explicit and implicit feedback features, which combines the IBPR model and the BiasSVD model, uses the BiasSVD model to extract explicit feedback features of users and items, and uses the IBPR model to extract implicit feedback features of users and items, reconstructs the user-item ranking score matrix according to the extracted implicit feedback features, reconstructs the user-item score matrix according to the extracted explicit feedback features, and fuses the reconstructed matrices, and the fused matrix is the final ranking score matrix, which fully utilizes the historical rating data and implicit feedback data in the data set to alleviate the cold start problem of the recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The framework diagram of the post-fusion personalized recommendation model based on explicit and implicit feedback features provided for the embodiment 1 of the present application is shown in the figure.

[0045] Figure 2 The pair-ordered set diagram comparison between the BPR model and the proposed IBPR model is shown in the figure.

[0046] Figure 3 The curve of the MAP / MRR index of the IBPR_SVD model on the public data set Movielens 100K with the value change of the compromise parameter alpha is shown in the figure.

[0047] Figure 4 The curve of the MAP / MRR index of the IBPR_SVD model on the public data set FilmTrust with the value change of the compromise parameter alpha is shown in the figure.

[0048] Figure 5 The schematic diagram of the attention mechanism module is shown in the figure.

[0049] Figure 6 Figure 2 is a curve of the MAP index of the IBPR_SVD model changing with the number of ratings;

[0050] Figure 7 Figure 3 is a curve of the MRR index of the IBPR_SVD model changing with the number of ratings;

[0051] Figure 8 Figure 4 is a flowchart of the post-fusion personalized recommendation method based on explicit and implicit feedback features provided in Embodiment 2 of the present application. DETAILED DESCRIPTION

[0052] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Embodiment 1

[0054] The present embodiment proposes a post-fusion personalized recommendation model based on explicit and implicit feedback features, which includes an explicit feature extraction module, an implicit feature extraction module and an overall feature extraction module, wherein the explicit feature extraction module and the implicit feature extraction module are connected with the overall feature extraction module. The explicit feature extraction module is used to receive data of a user feature matrix and an item feature matrix, extract explicit feedback features of users and items, and reconstruct a predicted rating matrix. The implicit feature extraction module is used to receive explicit rating data and implicit feedback data of users, extract implicit feedback features of users and items, and reconstruct a predicted ranking rating matrix. The overall feature extraction module is used to perform weighted summation on the predicted rating matrix and the predicted ranking rating matrix to obtain a final predicted ranking rating matrix.

[0055] Referring to FIG. 1, Figure 1 , the part marked BiasSVD algorithm in FIG. 1 is the explicit feature extraction module, the part marked IBPR algorithm in FIG. 1 is the implicit feature extraction module, and the right part of FIG. 1 is the overall feature extraction module. Figure 1 In FIG. 1, U is a user feature matrix in the BiasSVD algorithm, V is an item feature matrix in the BiasSVD algorithm, and the predicted rating matrix is Figure 1 Figure 1 In FIG. 1, U1 is a user feature matrix in the IBPR algorithm, V1 is an item feature matrix in the IBPR algorithm, and the predicted ranking rating matrix is

[0056] ​​In this embodiment, the explicit feature extraction module adopts BiasSVD model (Bias Singular Value Decomposition), the implicit feature extraction module adopts BPR model, and IBPR model can also be adopted, and the overall feature extraction module adopts IBPR_SVD model.

[0057] Based on the assumption that the user prefers the item with a high score value, the embodiment adds the definition of the model on the explicit rating data on the basis of the Bayesian personalized ranking model, and proposes an improved Bayesian personalized ranking IBPR (Improved Bayesian Personalized Ranking, IBPR) model to extract the implicit feedback features of the browsing user and the item. The model fully utilizes the existing browsing user feedback data and introduces the definition of negative feedback data, alleviates the problem of lack of negative feedback in the implicit feedback data, and improves the noise resistance of the model.

[0058] In order to alleviate the cold start problem, the embodiment adopts BiasSVD model to extract the explicit feedback features of the user and the item, reconstructs the user-item ranking score matrix according to the extracted implicit feedback features, reconstructs the user*item score matrix of the explicit feedback features, and fuses the reconstructed matrices, and proposes a post-fusion personalized recommendation model based on explicit and implicit feedback features. Since the model combines the IBPR model and the BiasSVD model, the model is simply referred to as IBPR_SVD (Improved Bayesian Personalized Ranking Singular Value Decomposition, IBPR_SVD).

[0059] From the definition of the BPR (Bayesian Personalized Ranking, BPR) model, it can be known that the model only extracts the features of the user and the item from the partial order set D S , and ignores the user-labeled project pairs. S The definition of D

[0060]

[0061] Wherein, the project i represents any one project labeled by the user u, the project j represents any one project not labeled by the user u. IV (u) represents a set of projects labeled by the user u, represents a set of projects not labeled by the user u. The triple (u, i, j) represents that the user u prefers the project i to the project j.

[0062] In browsing systems, a user typically rates several or even more browsing items simultaneously. The rating values ​​clearly indicate that users prefer items with higher ratings. Based on this, this embodiment proposes the IBPR model, which assumes users prefer items with higher ratings. Unlike the BPR model, the IBPR model proposed in this embodiment expands the pairwise data set input to the BPR model by adding a definition of the user's historical rating pairs. The expanded part includes the partially ordered set D. R The definition is as follows:

[0063] D R ={(u,i,j)|i∈Ⅳ(u)and j∈Ⅳ(u)and r ui >r uj} (2)

[0064] Where item i and item j represent two items that user u has marked, N(u) represents the set of items that user u has marked, and r ui and r uj Let represent user u's historical ratings for item i and item j, respectively. The triple (u, i, j) indicates that user u prefers item i more than item j. Therefore, the partially ordered set D of the IBPR model input can be represented as:

[0065] D = D S ∪D R (3)

[0066] The partially ordered set D has been expanded to include some pairs of users’ historical rating items.

[0067] The BPR model hierarchical dataset and the IBPR model hierarchical dataset in this embodiment are referenced in Appendix. Figure 2 .

[0068] Appendix Figure 2 This demonstrates how the BPR and IBPR models transform user feedback data into hierarchical partially ordered sets D, respectively. S The difference from D. (Appendix) Figure 2 The leftmost user-item rating matrix represents a set of user ratings for items. The scores reflect the user's level of liking; higher scores indicate greater user enjoyment of the item. Question marks indicate that a user has never rated the item. The BPR and IBPR models generate preference item pairs u:i> between two different items based on user labeling. u In a partially ordered set, a plus sign indicates that user u prefers item i to item j, while a minus sign indicates the opposite, meaning that user u prefers item j to item i. A question mark indicates that it is impossible to determine the relative preference of users for the two items.

[0069] From the appendixFigure 2 As can be seen, compared with the BPR model, the partial user historical rating items are added in the partial order set of the IBPR model. Taking the triple <u1, i2, i3> as an example, in the BPR model, the preference relationship of the user u1 to the item i2 and the item i3 cannot be judged. Since the rating records of the user u1 to the item i2 and the item i3 are 5 and 3 respectively, the IBPR model considers that the user u1 prefers i2 to i3. If the rating values of the user to two items are the same, the IBPR model cannot judge the preference relationship between the two items.

[0070] Based on the above analysis, the objective function of the IBPR model is:

[0071]

[0072] Wherein, σ(x) is a Sigmoid function, λ is a regularization parameter, Θ is a parameter set of the model, Θ = <b u , b i , p u , q i , q j}. is used to obtain the relationship between the user u and two items i and j, is defined as follows:

[0073]

[0074] Wherein, and respectively represent the ranking priority of the user u to the item i and the item j predicted by the model, b i represents the bias item of the item i, r max and r min respectively represent the maximum value and the minimum value of the user rating in the data set, p u represents the feature vector of the user u, q i represents the feature vector of the item i, represents the transpose of q i . The IBPR model is optimized and solved by the stochastic gradient descent method.

[0075] In this embodiment, the explicit feature extraction module adopts the BiasSVD model, that is, the BiasSVD model is used to extract the explicit feedback features of the user and the item. The rating prediction formula of any user u on the item i of the BiasSVD model adopts Wherein, μ is the rating value of the historical rating data of the training set, b i and b u respectively represent the bias item of the item i and the bias item of the user u, p uq i represents a feature vector of item i, represents q i .

[0076] The IBPR_SVD model is adopted in the overall feature extraction module of the embodiment, which combines the BiasSVD model and the IBPR model.

[0077] For the user u, the ranking score of any unlabeled item i predicted is which can be calculated by the following formula:

[0078]

[0079] wherein, represents the rating value of the user u on the item i predicted by the BiasSVD model, represents the ranking score of the user u on the item i predicted by the IBPR model. The value range of the compromise parameter a is [0, 1].

[0080] Since the value of the compromise parameter a directly affects the recommendation accuracy of the algorithm proposed in the embodiment, before verifying the performance of the IBPR_SVD algorithm, the optimal value of the compromise parameter a needs to be determined first. The greater the compromise parameter a, the greater the influence of the IBPR model on the IBPR_SVD algorithm, and the smaller the influence of the BiasSVD model. The smaller the compromise parameter a, the smaller the contribution of the IBPR model to the IBPR_SVD algorithm, and the greater the contribution of the BiasSVD model. When a = 0, the IBPR_SVD algorithm degenerates into the BiasSVD model, and when a = 1, the IBPR_SVD algorithm degenerates into the IBPR model. The specific value of the compromise parameter a is determined by experiments.

[0081] The most suitable value of the compromise parameter a is determined by experiments in the embodiment. The experiments are performed on two public data sets Movielens 100K and FilmTrust, the compromise parameter a is a variable, the value range is [0, 1], the step is 0.1, MAP and MRR are used as the evaluation indexes of the experiments, the iteration number is set to 1000, and the feature vector dimensions of the IBPR model and the BiasSVD model are both set to 10. The performance curves of the IBPR_SVD algorithm on the two data sets Movielens 100K and FilmTrust with the change of the compromise parameter a are shown in FIGS. 1 and 2 of the accompanying drawings. Figure 3 and FIGS. 3 and 4 of the accompanying drawings. Figure 4 It can be seen from FIGS. 1 and 2 of the accompanying drawings and FIGS. 3 and 4 of the accompanying drawings that the optimal value of the compromise parameter is 0.9. Figure 3 Figure 4

[0082] ​​Different from the prior art which takes all historical rating items as positive feedback data, the embodiment defines negative feedback data in the historical rating items; a compromise parameter is used to connect the collaborative filtering algorithm BiasSVD based on historical rating data and the collaborative filtering algorithm IBPR based on ranking rating, and the explicit features and implicit features of users and items are extracted jointly. The historical rating data and implicit feedback data in the data set are fully utilized, and the cold start problem of the recommendation system is alleviated.

[0083] Embodiment 2:

[0084] The embodiment provides a post-fusion personalized recommendation method based on explicit and implicit feedback features, referring to the accompanying drawings Figure 8 The post-fusion personalized recommendation method based on explicit and implicit feedback features fuses the BiasSVD model and the IBPR model, and includes the following steps:

[0085] Step one: obtaining explicit rating data and implicit feedback data of users from a user feedback database;

[0086] Step two: extracting explicit feedback features of users and items by using the BiasSVD model, learning a user feature matrix and an item feature matrix according to historical rating data of users, and reconstructing a predicted rating matrix through the learned feature matrix;

[0087] Step three: extracting implicit feedback features of users and items by using the IBPR model, constructing a pair-level item preference set of the IBPR model through explicit rating data and implicit feedback data of users, obtaining another group of user feature matrix and item feature matrix by training the IBPR model on the extended pair-level item set, predicting ranking ratings of all unlabeled items in the data set according to a ranking rating prediction formula, and further reconstructing a corresponding user-item predicted ranking rating matrix;

[0088] Step four: performing weighted summation on the predicted rating matrix obtained by the BiasSVD model and the predicted ranking rating matrix obtained by the IBPR model to obtain a final predicted ranking rating matrix, and the higher the ranking rating is, the greater the probability of obtaining a recommendation is;

[0089] Step five: ranking all predicted ranking ratings of users from high to low to obtain a recommendation list, and recommending the first four items with high ranking to users.

[0090] This embodiment of the post-fusion personalized recommendation method based on explicit and implicit feedback features combines the IBPR model and the BiasSVD model. The BiasSVD model is used as the explicit feature extraction module to extract explicit feedback features of users and items, reconstructing the predicted rating matrix. The IBPR model is used as the implicit feature extraction module to extract implicit feedback features of users and items, reconstructing the predicted ranking rating matrix. The IBPRSVD model is used as the overall feature extraction module to fuse the predicted rating matrix and the predicted ranking rating matrix to obtain the final predicted ranking rating matrix. The specific method can be referred to the processing procedure of the post-fusion personalized recommendation model based on explicit and implicit feedback features in Embodiment 1. This embodiment of the post-fusion personalized recommendation method based on explicit and implicit feedback features fully utilizes historical rating data and implicit feedback data in the dataset, alleviating the cold start problem of the recommendation system.

[0091] The workflow of this embodiment takes the following inputs: dataset T, IBPR_SVD model feature vector dimension f, learning rate γ, regularization parameter λ, number of iterations (interations), set of partial order pairs D, number of users m in the dataset, and number of items n; the output is the predicted ranking rating matrix. Initialize the preRmse parameter set in the BiasSVD and BPR models; start running until iteration stops.

[0092] To explore the most effective way to fuse BiasSVD and IBPR, this embodiment designs 13 different fusion methods for comparison. To more clearly illustrate these methods, this embodiment assumes that the user and item feature vectors of the BiasSVD model are U and V, respectively, and the user and item feature vectors of the IBPR model are P and Q, respectively. The 13 fusion methods are defined as follows:

[0093] 1) IBPR_SVD_AM. This embodiment designs an attention mechanism module to assign weights to BiasSVD and IBPR, as shown in the attached diagram. Figure 5 As shown. In the IBPR_SVD_AM method, the attention weights may differ for different users. (Appendix) Figure 5 The inputs a(u) and b(u) represent the influence of BiasSVD and IBPR on user u, respectively, and the output vector... This represents the user attention weight vector resulting from the influence of BiasSVD. These represent the user attention weight vectors resulting from the influence of IBPR. This embodiment designs a loss function to calculate the vectors. and

[0094]

[0095] wherein, denotes the training set, r ui denotes historical rating data, and λ is a regularization parameter. Gradient descent method is used to optimize the loss function.

[0096] 2) IBPR_SVD_MAX. This method takes the maximum value of BiasSVD and IBPR prediction results as the final prediction result.

[0097] 3) IBPR_SVD_MIN. This method takes the minimum value of BiasSVD and IBPR prediction results as the final prediction result.

[0098] 4) IBPR_SVD_MUL. This method takes the product of BiasSVD and IBPR prediction results as the final prediction result.

[0099] 5) IBPR_SVD_(U+P)(V+Q). In order to try more possibilities, this method recombines the feature sets of BiasSVD and IBPR models. This method first calculates the sum of user preference features and the sum of item attribute features of the two models, and then performs rating prediction for unrated items.

[0100] 6) IBPR_SVD(UQ+VP). This method recombines the features of the two models, exchanges the item features of BiasSVD and IBPR models, and calculates the corresponding dot products.

[0101] 7) IBPR_SVD_UQ. This method extracts the user features of BiasSVD and the item features of IBPR model for recommendation.

[0102] 8) IBPR_SVD_VP. This method extracts the item features of BiasSVD and the user features of IBPR model for recommendation.

[0103] 9) IBPR_SVD(U+P)V. This method uses the sum of user features of the two models and the item features of BiasSVD model for recommendation.

[0104] 10) IBPR_SVD(U+P)Q. This method uses the sum of user features of the two models and the item features of IBPR model for recommendation.

[0105] 11) IBPR_SVD_U(V+Q). This method uses the user features of BiasSVD model and the item features of the two models for recommendation.

[0106] 12) IBPR_SVD_P(V+Q). This method uses the user features of IBPR model and the item features of two models to make recommendations.

[0107] 13) IBPR_SVD. This method is the method proposed in this embodiment, which uses the compromise parameter a to fuse BiasSVD and IBPR models, and the compromise parameter is set to 0.9.

[0108] According to the above description, the performance of the 13 fusion methods in this embodiment is tested on two public data sets Movielens 100K and FilmTrust, the evaluation indexes are selected as Precision@3, Recall@3, Precision@5, Recall@5, MAP and MRR, and the experimental results are shown in Table 1. From the experimental results, it can be seen that the fusion method proposed in this embodiment is significantly better than other fusion methods.

[0109] Table 1

[0110]

[0111]

[0112] In order to illustrate the effectiveness of the method IBPR_SVD proposed in this embodiment, a large number of experiments are carried out on five public data sets, and the data sets are Movielens 100K, Movielens 1M, FilmTrust, Ciao and Hetrec-movielens-2k. Four relatively novel methods are compared to verify the effectiveness of the IBPR_SVD method. The indexes used in the experiment are Precision@3, Recall@3, Precision@5, Recall@5, MAP and MRR, and the experimental results are shown in Table 2. The experimental results show that the performance of the proposed IBPR_SVD is significantly better than other methods, which proves again that the method proposed in this embodiment is effective.

[0113] Table 2

[0114]

[0115]

[0116]

[0117] The models compared in this embodiment have all been published in top journals in the field of information recommendation, including Knowledge-Based Systems (KBS) and Information Sciences (Inf.Sci). The models compared are: 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).

[0118] This embodiment explores the performance of the IBPR_SVD method under user cold start conditions. The new user cold start dataset used in this embodiment was artificially generated from the FilmTrust dataset. Specifically, the cold start dataset was generated by varying the number of user ratings, with the number of user ratings ranging from [3, 19] and a step size of 2. The number of user ratings varies depending on the dataset. This embodiment uses the SPR method for comparison and MAP and MRR as evaluation metrics. Experimental results are attached. Figure 6 and attached Figure 7 As shown, from the appendix Figure 6 and attached Figure 7 As can be seen from this, the IBPR_SVD method proposed in this embodiment can alleviate the cold start problem to a certain extent.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. All content that does not depart from the technical solution of the present invention should be included within the protection scope of the present invention.

Claims

1. A post-fusion personalized recommendation system based on explicit and implicit feedback features, characterized in that, It includes an explicit feature extraction module, an implicit feature extraction module, and an overall feature extraction module, wherein the explicit feature extraction module and the implicit feature extraction module are respectively connected to the overall feature extraction module; The explicit feature extraction module is used to receive data from the user feature matrix and the project feature matrix, extract explicit feedback features of users and projects, and reconstruct the predicted rating matrix. The implicit feature extraction module is used to receive explicit rating data and implicit feedback data from users, extract implicit feedback features of users and items, and reconstruct the predicted ranking rating matrix. The overall feature extraction module is used to perform a weighted summation of the predicted score matrix and the predicted ranking score matrix to obtain the final predicted ranking score matrix. The implicit feature extraction module adopts the IBPR model; The IBPR model is used to expand the pairwise data set input to the BPR model, adding the user's definition on historical rating item pairs, and the partially ordered set of the expanded part. The definition is as follows: Among them, the project and projects Indicates user Two of the items were marked. Indicates user A collection of items that have been previously marked. and Representing users respectively In the project and projects Historical rating values, triplet Indicates relative to the project ,user More preference for projects Therefore, the partially ordered set of input Represented as: The partially ordered set The system has added some user history rating items; The overall feature extraction module adopts the IBPR_SVD model; The IBPR_SVD model integrates the BiasSVD model and the IBPR model; For users Any unlabeled item predicted Sorting score Calculate using the following formula: in, This represents the user predicted by the BiasSVD model. In the project The rating on the screen, This indicates the user predicted by the IBPR model. In the project The ranking score on the top As a compromise parameter, The range of values ​​is .

2. The post-fusion personalized recommendation system based on explicit and implicit feedback features according to claim 1, characterized in that, The implicit feature extraction module uses the BPR model; The BPR model is used to extract from partially ordered sets. Extract user and project features; The partially ordered set The definition is as follows: Among them, the project Indicates user Any item that has been marked, item Indicates user Any unmarked item, Indicates user A collection of items that have been previously marked. Indicates user Unlabeled collection of items, triples Indicates relative to the project ,user More preference for projects .

3. The post-fusion personalized recommendation system based on explicit and implicit feedback features according to claim 1, characterized in that, The objective function of the IBPR model is: in, For the Sigmoid function, For regularization parameters, For the parameter set of the IBPR model, , Used to acquire users and two projects and The relationship between them The definition is as follows: in, and These represent the users predicted by the IBPR model. For the project and projects The sorting priority, , Indicates project The bias term, and These represent the maximum and minimum user ratings in the dataset, respectively. Indicates user eigenvectors, Indicates project eigenvectors, express The transpose of .

4. The post-fusion personalized recommendation system based on explicit and implicit feedback features according to claim 3, characterized in that, The explicit feature extraction module uses the BiasSVD model; Any user of the BiasSVD model In the project The scoring prediction formula uses ,in, The rating values ​​are the historical rating data of the training set. and Each represents a project Bias terms and users The bias term, Indicates user eigenvectors, Indicates project eigenvectors, express The transpose of .

5. The post-fusion personalized recommendation system based on explicit and implicit feedback features according to claim 1, characterized in that, The compromise parameters The optimal value is 0.

9.

6. A post-fusion personalized recommendation method based on explicit and implicit feedback features, characterized in that, Includes the following steps: Step 1: Obtain explicit rating data and implicit feedback data from the user feedback database; Step 2: Use the BiasSVD model to extract explicit feedback features of users and items, learn the user feature matrix and item feature matrix based on the user's historical rating data, and reconstruct the predicted rating matrix using the learned feature matrix. Step 3: Use the IBPR model to extract implicit feedback features of users and items. Construct the pairwise item preference set of the IBPR model using explicit rating data and implicit feedback data of users. Train the IBPR model on the expanded pairwise item set to obtain another set of user feature matrices and item feature matrices. Predict the ranking ratings of all unlabeled items in the dataset according to the ranking rating prediction formula, and then reconstruct the corresponding user-item predicted ranking rating matrix. Step 4: Take the weighted sum of the predicted rating matrix obtained by the BiasSVD model and the predicted ranking rating matrix obtained by the IBPR model to obtain the final predicted ranking rating matrix. The higher the ranking rating, the greater the probability of receiving a recommendation. Step 5: Sort all the ranking scores predicted by the user from high to low to obtain the recommendation list, and then select the top-ranked ones. Recommend a project to the user; The weighted summation is performed on the user Any unlabeled item predicted Sorting score This is achieved through the sorting score. The expression is: in, This represents the user predicted by the BiasSVD model. In the project The rating on the screen, This indicates the user predicted by the IBPR model. In the project The ranking score on the top As a compromise parameter, The range of values ​​is ; Negative feedback data is defined in the historical rating items. A trade-off parameter is used to connect the collaborative filtering algorithm BiasSVD based on historical rating data and the collaborative filtering algorithm IBPR based on ranking ratings to jointly extract the explicit and implicit features of the users and items. The IBPR model is an extension of the pairwise data set input to the BPR model, adding user definitions on historical rating item pairs, and a partially ordered set in the extended part. The definition is as follows: Among them, the project and projects Indicates user Two of the items were marked. Indicates user A collection of items that have been previously marked. and Representing users respectively In the project and projects Historical rating values, triplet Indicates relative to the project ,user More preference for projects Therefore, the partially ordered set of input Represented as: The partially ordered set The system has added some user history rating items.

Images