A multi-domain content recommendation method and system that fuses novelty

Abstract

The application provides a multi-field content recommendation method and system with novelty fusion, and the application optimizes a user-item rating matrix based on a BiasSVD model, adds a novelty factor to predict the user rating matrix, and alleviates problems such as unsatisfactory recommendation effect caused by data sparseness in a traditional model; similarity is calculated based on user feature information, cold start problems caused by excessive dependence of a traditional recommendation algorithm on historical data are solved, the recommendation result is closer to the real rating, and the recommendation accuracy is improved; the novelty relationship between users is introduced, and the novelty of the recommendation result of a traditional collaborative filtering recommendation algorithm is improved. The method disclosed in the application can alleviate the problem of sparse rating data, and can improve the novelty of the recommendation result, so that the recommendation result is not too single, thereby avoiding user aesthetic fatigue.

Description

Technical Field

[0001] This invention relates to the field of recommendation systems in data mining, and in particular to a multi-domain content recommendation method and system that integrates novelty. Background Technology

[0002] With the development and changes of the internet age, much information can be easily and conveniently obtained online. However, people's needs and preferences in life and entertainment are often not so clear; they mostly search for interesting information aimlessly among the messages they receive. At the same time, the sheer volume of information on the internet can be overwhelming. This new dilemma is known as information overload. How can consumers find truly interesting information amidst this deluge, and how can those who produce information showcase their work amidst the overwhelming amount of data to gain wider acceptance? These are challenges facing people today. In this context, recommendation systems have emerged, helping users obtain resources that match their interests and needs from massive amounts of data at a lower cost.

[0003] The overall effectiveness of a recommendation system relies on its algorithms. Among these, collaborative filtering is one of the most successful techniques in the field. It is divided into user-based collaborative filtering and item-based collaborative filtering. Traditional user-based collaborative filtering mainly finds user groups with similar interests to the target user, predicts the target user's preferences based on their favorite items, and generates recommendations. Item-based collaborative filtering compares each item rated by the target user with all other unrated items, finds the most similar neighboring items, predicts ratings for highly similar neighboring items, and selects the top N highest-rated items to recommend to the target user. Although collaborative filtering technology has shown great advantages in information filtering, it has also encountered some problems in different applications. For example, how to recommend items to newly registered users or newly added items to satisfy them, and how to distribute new items to users who like them? Secondly, the user-item rating matrix is ​​extremely sparse, resulting in low accuracy in calculating the nearest neighbors of users or items, thus drastically reducing the recommendation quality of the system. The most significant issue is the inability to differentiate and select items with high novelty for different users. The cold start problem, data sparsity problem, and novelty problem of recommendation results have not yet been well resolved. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention discloses a multi-domain content recommendation method and system that integrates novelty. This multi-domain recommendation method combines matrix factorization with novelty integration with system filtering incorporating user features. This method can provide more novel recommendations to different users, reduce the impact of data sparsity on collaborative filtering algorithms, and improve the accuracy of recommendation algorithms.

[0005] To achieve the above objectives, this invention provides a multi-domain content recommendation method that integrates novelty, which mainly includes the following steps:

[0006] S1, Process the user-item rating matrix, and process the user's historical item rating data into an m*n user-item rating matrix R, where m represents the number of users and n represents the number of items;

[0007] S2, decompose the user-item rating matrix into the product of two lower-dimensional rating matrices P and Q;

[0008] S3, based on the BiasSVD scoring prediction formula The loss function for building the model:

[0009]

[0010] Wherein, λ(||p u || 2 +||q i || 2 +||b u || 2 +||b i || 2 ) is a regularization term to prevent overfitting, where the elements r in the matrix ui p represents the actual rating of user u for item i. u This represents a row of data in matrix P, q u This represents a column of data in the Q matrix, b u and b i These represent user biases and project biases, respectively.

[0011] S4. Use gradient descent to optimize the loss function. Take the partial derivatives with respect to the parameters and calculate the update formula along the direction of the fastest gradient descent.

[0012] S5. Calculate the novelty factor using the user's historical project information, and incorporate the P and Q matrices obtained in step S2 to predict the rating of unrated projects.

[0013] S6 incorporates user characteristics such as gender, age, and occupation into the similarity calculation, selects user feature similarity and Pearson similarity with different proportions, and predicts the rating of unrated items with the optimal similarity.

[0014] S7. Select the top N items with the highest ratings obtained from the previous steps and recommend them to the user.

[0015] Furthermore, in step S2, the user item rating matrix is ​​decomposed into two low-dimensional matrices P and Q, calculated using the following formula:

[0016]

[0017] Where R is an m*n user-item rating matrix, where m represents each user and n represents n items, P is an m*k matrix, representing the matrix composed of m users and k latent features, and Q is a k*n matrix, representing the matrix composed of n items and k latent features;

[0018] The element values ​​in matrix R are:

[0019] Where, p u This represents a row of data in matrix P. This represents a column of data in matrix Q. This represents user u's predicted rating for item i;

[0020] Furthermore, the update formula in step S4 is as follows:

[0021]

[0022]

[0023]

[0024]

[0025] Where α is the learning rate and λ is the regularization parameter, the score can be predicted using the scoring formula by training the parameters on the training set.

[0026] Furthermore, the formula for calculating the novelty factor in step S5 is as follows:

[0027] N = ηF + (1-η)T

[0028] in,

[0029] F represents the proportion of the total number of times a user visits a project to the number of visits to similar projects.

[0030] M u,i To measure the number of times a certain type of project is accessed.

[0031] M U Total number of times the project was accessed;

[0032] T represents the time interval between the user's last visit to a similar project and this visit.

[0033] t present The current access time.

[0034] t last This is the last access time.

[0035] Furthermore, step S5 predicts the score for unrated items, and the calculation formula is as follows:

[0036]

[0037] Where N is the novelty factor and β is a constant weight parameter used to control the impact of novelty on the recommendation results, preventing the accuracy from decreasing as the novelty of the recommendation results increases.

[0038] Furthermore, the similarity calculation formula in step S6 is as follows:

[0039]

[0040] Where, sim u User feature similarity is calculated using the following formula:

[0041]

[0042] in,

[0043] In the above formula, u i1 u i2 and u i3 User i represents the quantified value of gender, age, and one of the occupational attributes, respectively; u j1 ,u j2 ,u j3 User j represents the quantified values ​​of gender, age, and one of the occupational attributes, respectively.

[0044] Where Pearson is the Pearson similarity, calculated using the following formula:

[0045]

[0046] Among them, R u1，i R represents user u1's rating of item i. u2，i This represents user u2's rating of item i. This represents the average score of the items rated by user u1. The average score of the items rated by user u2;

[0047] Furthermore, we calculate the similarity of user characteristics such as gender, age, and occupation. First, we quantify these characteristics. For gender, we define male as 1 and female as 0. For age, we divide it into Range = {1, 2, 3, 4, 5, 6}, where each number represents an age range: 1: 0-14, 2: 15-23, 3: 24-33, 4: 34-45, 5: 45-60, 6: over 60 years old. For occupation, it is difficult to measure the differences between different occupations, so we use each number in {0, 1, 2, 3, 4, 5...20} to represent one of the 21 common occupations, where u... i3 with u j3 The following relationships exist:

[0048]

[0049] Furthermore, the formula for calculating the score of unrated items in step S6 is as follows:

[0050]

[0051] Where r vi Sim(u,v) represents the ratings of users similar to the target user for the target item, and Sim(u,v) represents the total similarity between the user and the target user.

[0052] Meanwhile, this invention also provides a multi-domain content recommendation system that integrates novelty. The system includes a matrix factorization module, a novelty calculation module, a similarity calculation module, and a prediction module. The matrix factorization module decomposes the user-item rating matrix into the product of two lower-dimensional rating matrices P and Q. The novelty calculation module calculates a novelty factor using the user's historical item information and incorporates it into the P and Q matrices in the matrix factorization module to predict the rating of unrated items. The similarity module incorporates the user's gender, age, occupation, and other features into the similarity calculation, selects user feature similarity and Pearson similarity with different proportions, and predicts the rating of unrated items with the optimal similarity. Finally, the prediction module recommends the top N items with the highest ratings to the user.

[0053] The advantages and beneficial effects of this invention are as follows:

[0054] (1) The user item rating matrix is ​​optimized and decomposed based on the BiasSVD model, and the novelty factor is added to predict the user rating matrix, which alleviates the problem of unsatisfactory recommendation effect caused by data sparsity in traditional models.

[0055] (2) Calculate similarity based on user feature information to solve the cold start problem caused by the over-reliance on historical data in traditional recommendation algorithms. This also makes the recommendation results closer to the real ratings and improves the accuracy of recommendations.

[0056] (3) Introduce novelty relationships between users to improve the novelty of recommendation results from traditional collaborative filtering recommendation algorithms. Attached Figure Description

[0057] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of a multi-domain content recommendation method that integrates novelty according to the present invention;

[0059] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0060] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0061] Main English meanings:

[0062] BiasSVD (Bias Singular Value Decomposition): SVD decomposition with bias terms. The invention will be further described below with reference to the accompanying drawings.

[0063] To achieve the above objectives, this invention provides a multi-domain content recommendation method that integrates novelty, which mainly includes the following steps:

[0064] S1, Process the user-item rating matrix, and process the user's historical item rating data into an m*n user-item rating matrix R, where m represents the number of users and n represents the number of items;

[0065] S2, decompose the user-item rating matrix into the product of two lower-dimensional rating matrices P and Q;

[0066] S3, based on the BiasSVD scoring prediction formula The loss function for building the model:

[0067]

[0068] Wherein, λ(||p u || 2 +||q i || 2 +||b u || 2 +||b i ||2 ) is a regularization term to prevent overfitting, where the elements r in the matrix ui p represents the actual rating of user u for item i. u This represents a row of data in matrix P, q u This represents a column of data in the Q matrix, b u and b i These represent user biases and project biases, respectively.

[0069] S4. Use gradient descent to optimize the loss function. Take the partial derivatives with respect to the parameters and calculate the update formula along the direction of the fastest gradient descent.

[0070] S5. Calculate the novelty factor using the user's historical project information, and incorporate the P and Q matrices obtained in step S2 to predict the rating of unrated projects.

[0071] S6 incorporates user characteristics such as gender, age, and occupation into the similarity calculation, selects user feature similarity and Pearson similarity with different proportions, and predicts the rating of unrated items with the optimal similarity.

[0072] S7. Select the top N items with the highest ratings obtained from the previous steps and recommend them to the user.

[0073] This embodiment uses the film industry as an example to illustrate the solution:

[0074] In step S1, the user-movie rating matrix has columns named "user" and rows named "movie". There are m users and n items. There are unrated items in the rating data, which are temporarily represented by 0. In addition, the number of items rated by users in the user rating data is much smaller than the total number of items, so the matrix here is a sparse matrix.

[0075] In step S2, the user item rating matrix is ​​decomposed into two low-dimensional matrices P and Q, and the calculation formula is as follows:

[0076]

[0077] Where R is an m*n user-movie rating matrix, where m represents each user and n represents n movies, P is an m*k matrix, representing the matrix composed of m users and k latent features, and Q is a k*n matrix, representing the matrix composed of n items and k latent features;

[0078] The element values ​​in matrix R are:

[0079] Where, p u This represents a row of data in matrix P. This represents a column of data in matrix Q. This represents user u's predicted rating for item i;

[0080] The specific steps for decomposing the user movie rating matrix into two low-dimensional matrices are as follows:

[0081] 1) First, randomly generate an m x k ​​matrix and a k x n matrix, where k is the number of latent factors. Select an appropriate value based on the training results. The elements in both matrices are random numbers between 0 and 1.

[0082] 2) Based on the predicted score And actual rating r ui , through formula Calculate the error value;

[0083] 3) The loss function is minimized using gradient descent:

[0084] 4)

[0085] 5) By analyzing the characteristic variable p u ,q i Taking the partial derivatives of each, and applying them along the direction of fastest gradient descent, we can obtain their update formulas:

[0086] 6)p u =p u +α(( ui -∑p u q i ) i -p u )

[0087] 7)q i = i +α(( ui -∑p u q i ) u -q i )

[0088] 8) By updating the above parameters, the error is continuously reduced. After a certain number of iterations, training is stopped when the error is very small. The parameters obtained from training can be used to predict unrated items and fill the sparse rating matrix.

[0089] By incorporating user bias and project bias into the previous rating formula, we can correct for user and project biases, as different users have varying levels of rigor in their ratings and different projects are viewed with varying degrees of popularity. Therefore, the rating prediction formula becomes:

[0090]

[0091] Where u is the global average score, bu For user paranoia, b i For project bias, the loss function becomes:

[0092]

[0093] The update formulas for each parameter can be obtained using the same method described above:

[0094]

[0095]

[0096]

[0097]

[0098] Similarly, the trained parameters can be used to predict unknown ratings and fill in the sparse rating matrix;

[0099] In step S5, the novelty factor is calculated mainly from two aspects:

[0100] First, the proportion of similar items in the total number of user visits to a category. A higher visit rate indicates lower novelty for that category, while a lower visit frequency indicates higher novelty. The calculation formula is as follows:

[0101] Where F represents the proportion of the total number of times a user visits a project that is of the same type.

[0102] M u,i To measure the number of times a certain type of project is accessed.

[0103] M U Total number of times the project was accessed;

[0104] The second factor is the time interval between the user's last visit to a similar project and this visit. The longer the interval, the higher the novelty level. The calculation formula is as follows:

[0105] T represents the time interval between the user's last visit to a similar project and this visit.

[0106] t present The current access time.

[0107] t last This is the last access time.

[0108] Combining these two aspects, we obtain the overall novelty formula as follows:

[0109] N = ηF + (1-η)T

[0110] Furthermore, the novelty and matrix factorization scoring formulas are combined to predict the ratings of unrated movies. The calculation formula is as follows:

[0111]

[0112] Where N is the novelty factor and β is a constant weight parameter used to control the impact of novelty on the recommendation results, preventing the accuracy from decreasing as the novelty of the recommendation results increases.

[0113] Furthermore, the similarity calculation formula in step S6 is as follows:

[0114]

[0115] Where, sim u User feature similarity is calculated using the following formula:

[0116]

[0117] in,

[0118] In the above formula, u i1 u i2 and u i3 User i represents the quantified value of gender, age, and one of the occupational attributes, respectively; u j1 ,u j2 ,u j3 User j represents the quantified values ​​of gender, age, and one of the occupational attributes, respectively.

[0119] Where Pearson is the Pearson similarity, calculated using the following formula:

[0120]

[0121] Among them, R u1,i R represents user u1's rating of item i. u2，i This represents user u2's rating of item i. This represents the average score of the items rated by user u1. The average score of the items rated by user u2;

[0122] Furthermore, we calculate the similarity of user characteristics such as gender, age, and occupation. First, we quantify these characteristics. For gender, we define male as 1 and female as 0. For age, we divide it into Range = {1, 2, 3, 4, 5, 6}, where each number represents an age range: 1: 0-14, 2: 15-23, 3: 24-33, 4: 34-45, 5: 45-60, 6: over 60 years old. For occupation, it is difficult to measure the differences between different occupations, so we use each number in {0, 1, 2, 3, 4, 5...20} to represent one of the 21 common occupations, where u... i3 with u j3 The following relationships exist:

[0123]

[0124] Based on the quantitative values ​​corresponding to the three user characteristics of gender, age, and occupation, we can use the above formula to calculate the similarity between two users.

[0125] Furthermore, based on the overall similarity calculation formula, we can calculate the similarity between pairs of users. We select the top n users with the highest similarity to the target user, and use their rating data to predict the target user's rating for the target item. The prediction formula is as follows:

[0126]

[0127] Where r vi Sim(u,v) represents the ratings of users similar to the target user for the target item, and Sim(u,v) represents the total similarity between the user and the target user.

[0128] Meanwhile, this invention also provides a multi-domain content recommendation system that integrates novelty. The system includes a matrix factorization module, a novelty calculation module, a similarity calculation module, and a prediction module. The matrix factorization module decomposes the user-item rating matrix into the product of two lower-dimensional rating matrices P and Q. The novelty calculation module calculates a novelty factor using the user's historical item information and incorporates it into the P and Q matrices in the matrix factorization module to predict the rating of unrated items. The similarity module incorporates the user's gender, age, occupation, and other features into the similarity calculation, selects user feature similarity and Pearson similarity with different proportions, and predicts the rating of unrated items with the optimal similarity. Finally, the prediction module recommends the top N items with the highest ratings to the user.

[0129] This invention utilizes a matrix factorization algorithm to fill the rating matrix, alleviating the data sparsity problem in movie recommendation systems. Furthermore, it incorporates a novelty factor into the rating, which can appropriately improve the novelty of recommendations while ensuring recommendation accuracy, preventing overly simplistic recommendations that could lead to user fatigue.

[0130] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

Claims

1. A multi-domain content recommendation method integrating novelty, characterized in that, This method mainly includes the following steps: S1. Establish a user item rating matrix by processing the user's historical item rating data into an m*n user item rating matrix R, where m represents the number of users and n represents the number of items. S2, decompose the user item rating matrix into the product of two lower-dimensional rating matrices P and Q; S3, based on the BiasSVD scoring prediction formula The loss function for building the model: ； in, This is a regularization term to prevent overfitting, where the elements in the matrix... p represents the actual rating of user u for item i. u This represents a row of data in matrix P, q i This represents a column of data in the Q matrix, b u and b i These represent user biases and project biases, respectively. S4. The loss function is optimized using gradient descent. The partial derivatives with respect to the parameters are calculated and the update formula is obtained by calculating the partial derivatives with respect to the parameters along the direction of the fastest gradient descent. S5. Calculate the novelty factor using the user's historical project information, and incorporate the P and Q matrices obtained in step S2 to predict the rating of unrated projects. S6 incorporates the user's gender, age, and occupation characteristics into the similarity calculation, selects user feature similarity and Pearson similarity with different proportions, and uses the optimal similarity to predict the rating of unrated items. S7. Select the top few items with the highest scores obtained in step S6 and recommend them to the user. The formula for calculating the novelty factor in step S5 is as follows: ; in, , ; F represents the proportion of the total number of times a user visits a project to the number of visits to similar projects. To measure the number of times a certain type of project is accessed. Total number of times the project was accessed; T represents the time interval between the user's last visit to a similar project and this visit. This is the current access time; The last access time; The similarity calculation formula in step S6 is as follows: ; in, User feature similarity is calculated using the following formula: ； in, ; In the above formula , and These represent the quantified values ​​of user i's gender, age, and one of its occupational attributes, respectively. , , These represent the quantified values ​​of user j's gender, age, and one of his / her occupational attributes, respectively. Where Pearson is the Pearson similarity, calculated using the following formula: ； in, This represents user u1's rating of item i. This represents user u2's rating of item i. This represents the average score of the items rated by user u1. This represents the average score of the items rated by user u2.

2. The multi-domain content recommendation method integrating novelty as described in claim 1, characterized in that, In step S2, the user item rating matrix is ​​decomposed into two low-dimensional matrices P and Q, and the calculation formula is as follows: ； Where R is an m*n user item rating matrix, where m represents each user and n represents n items, P is an m*k matrix, representing the matrix composed of m users and k latent features, and Q is a k*n matrix, representing the matrix composed of n items and k latent features. The element values ​​in matrix R are: ; Where, p u This represents a row of data in matrix P. This represents a column of data in matrix Q. This represents user u's predicted rating for item i.

3. The multi-domain content recommendation method integrating novelty according to claim 1, characterized in that, The update formula for step S4 is as follows: ; ; ; ; Where α is the learning rate and λ is the regularization parameter, the parameters trained on the training set are used to predict the score using the score prediction formula.

4. The multi-domain content recommendation method integrating novelty according to claim 1, characterized in that, Step S5 predicts the score for unrated items, and its calculation formula is as follows: ; Where N is the novelty factor and β is a constant weight parameter used to control the impact of novelty on the recommendation results, preventing the accuracy from decreasing as the novelty of the recommendation results increases.

5. The multi-domain content recommendation method integrating novelty according to claim 1, characterized in that, To calculate the similarity of users' gender, age, and occupation characteristics, we first quantify them. For gender, we define male as 1 and female as 0. For age, we divide it into Range = {1, 2, 3, 4, 5, 6}, where each number represents an age range: 1: 0-14, 2: 15-23, 3: 24-33, 4: 34-45, 5: 45-60, 6: over 60 years old. For occupation, it is difficult to measure the differences between different occupations, so we use each number in {0, 1, 2, 3, 4, 5...20} to represent one of the 21 common occupations, where u... i3 with u j3 The following relationships exist: ; 6. The multi-domain content recommendation method integrating novelty according to claim 2, characterized in that, The formula for calculating the score of unrated items in step S6 is as follows: ; in Sim(u,v) represents the ratings of users similar to the target user for the target item, and Sim(u,v) represents the total similarity between the user and the target user.

7. A multi-domain content recommendation system integrating novelty, characterized in that, The system includes a matrix factorization module, a novelty calculation module, a similarity calculation module, and a prediction module. The matrix factorization module decomposes the user item rating matrix into the product of two lower-dimensional rating matrices P and Q. The novelty calculation module calculates the novelty factor using the user's historical item information and incorporates it into the P and Q matrices in the matrix factorization module to predict the rating of unrated items. The similarity module incorporates the user's gender, age, and occupation characteristics into the similarity calculation, selects user feature similarity and Pearson similarity with different proportions, and predicts the rating of unrated items with the optimal similarity. Finally, the prediction module recommends the top few items with the highest ratings to the user. The formula for calculating the novelty factor is: ; in, , ; F represents the proportion of the total number of times a user visits a project to the number of visits to similar projects. To measure the number of times a certain type of project is accessed. Total number of times the project was accessed; T represents the time interval between the user's last visit to a similar project and this visit. This is the current access time; The last access time; The formula for calculating similarity is: ; in, User feature similarity is calculated using the following formula: ； in, ; In the above formula , and These represent the quantified values ​​of user i's gender, age, and one of its occupational attributes, respectively. , , These represent the quantified values ​​of user j's gender, age, and one of his / her occupational attributes, respectively. Where Pearson is the Pearson similarity, calculated using the following formula: ； in, This represents user u1's rating of item i. This represents user u2's rating of item i. This represents the average score of the items rated by user u1. This represents the average score of the items rated by user u2.

Images