Personalized intelligent recommendation method and system for cultural and tourism attractions based on federated learning

Abstract

The present invention relates to a personalized intelligent recommendation method for cultural and tourist attractions based on federated learning, comprising the following steps: defining the user's natural feature vector nature and the feature vector of interest hobby , and then consider the similarity between users by combining natural features and interest features sim , get the target user u Similar user clusters; will be u The attraction ratings given by all users in similar user clusters are integrated into the user-attraction rating matrix S； Using mass correction factors correction i First S Make corrections and then perform matrix decomposition prediction to get the target user u The first n desired attractions are selected, and then the order of the recommendation results is modified in combination with implicit feedback data to obtain the final recommendation result. The present invention can not only achieve accurate recommendations that meet personalized needs, but also ensure the privacy and security of the recommendation process.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and more particularly to a method and system for intelligently recommending personalized cultural and tourist attractions based on federated learning. Background Art

[0002] With the explosive growth of tourism resource information and the rapid development of deep learning technology, various recommendation algorithms have performed well in the field of tourism recommendation systems. Traditional recommendation algorithms focus on personalized recommendations, but most ignore issues such as the leakage of user privacy data and significant cold start in the process. Federated learning, on the other hand, is a distributed machine learning technology that focuses on achieving a balance between privacy protection and data sharing and computing. It enables joint training of global models without the need for raw data transfer when multiple data sources participate in model training, thus achieving the goal of keeping the raw data local. Compared to various traditional methods in the field of cultural and tourism recommendations, the personalized intelligent recommendation method for cultural and tourism attractions based on federated learning can ensure that users have a good personalized recommendation experience, alleviate cold start issues, and simultaneously achieve the security of user privacy information.

[0003] At present, federated learning technology has been widely and deeply used in various fields of deep learning. However, in many recommendation systems and methods in the cultural and tourism field, user privacy protection has always been a neglected link. How to ensure the diversity and personalization of recommendations while still taking into account user privacy data has become an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a personalized cultural and tourism attraction intelligent recommendation method and system based on federated learning, aiming to solve the above problems.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A personalized intelligent recommendation method for cultural and tourist attractions based on federated learning includes the following steps:

[0007] Define the user's natural feature vector nature and interest feature vector hobby, then consider the similarity sim between users by combining the natural features and interest features to obtain a user cluster similar to the target user u;

[0008] Integrate the attraction ratings given by all users in the user cluster similar to the target user u into the user-attraction rating matrix S;

[0009] Use quality correction factor correction iFirst, S is corrected, and then matrix decomposition prediction is performed to obtain the top n desired attractions of the target user u. Then, combined with implicit feedback data, the order of the recommendation results is corrected to obtain the final recommendation result.

[0010] Furthermore, the user's natural feature vector nature and interest feature vector hobby are defined as follows:

[0011] Define user age preference vector age=(z1,z2,z3,z4) T ,z i ∈(0,1), and divide all users into two groups according to gender Where M represents the male group and F represents the female group, and the quantification rules are preset as follows:

[0012]

[0013] The user's nature vector is calculated based on the quantization rule, where z1'=z1*1,z2'=z2*2,z3'=z3*3,z4'=z4*4. When user u is male, sex u = 0, when user u is female, sex u =1;

[0014] Divide attractions into four categories: cultural, consumer, internet celebrity, and environmental; and construct the user interest vector hobby = (x1, x2, x3, x4) T ,x i ∈(0,1), where x1, x2, x3, and x4 represent the user's preference for cultural attractions, consumer attractions, internet celebrity attractions, and environmental attractions, respectively. The sum of x1, x2, x3, and x4 should be 1.

[0015] Furthermore, the similarity sim between users is considered by combining natural features and interest features to obtain a user cluster similar to the target user u, as follows:

[0016] Step A1: The inter-user similarity sim is obtained by the user interest similarity sim1 and the user natural similarity sim2. First calculate sim1:

[0017] It is meaningless to use the Pearson correlation coefficient to standardize the user interest vector in advance, so consider using the cosine similarity method to calculate the similarity of user interests:

[0018]

[0019] Step A2: Calculate sim2 again, based on the Euclidean distance to reflect the differences between different users. The Euclidean distance formula between user u and user v is as follows:

[0020]

[0021] The larger the Euclidean distance, the greater the difference between user u and user v, and the smaller the similarity; the smaller the Euclidean distance, the smaller the difference between user u and user v, and the greater the similarity; therefore, the formula for sim2 is as follows:

[0022]

[0023] Finally get sim:

[0024] sim(u, υ)=(1-α)sim1(u, υ)+αsim2(u, υ).

[0025] Furthermore, the user-attraction rating matrix S is corrected and matrix decomposition is predicted. The specific method is as follows:

[0026] For the user-attraction rating matrix, it is expected to decompose the matrix S into: S m×n =p m×k ×q k×n ;

[0027] Assume that user u’s rating of scenic spot i is S u,i , after matrix decomposition, the feature vector of user u mapped to the k-dimensional space is p u , the characteristic vector of scenic spot n is q i , the predicted score is: Introducing the regularization factor λ, the loss function is:

[0028]

[0029] Use BP neural network to train and minimize the loss function. Let the time set of the first n-1 iterations be T = {t0, t1, ..., t n-1}, then the learning rate γ for the nth iteration i (t i )for:

[0030]

[0031] When the loss function reaches the minimum, p and q are obtained. Multiplying p and q together can obtain a denser user-attraction rating matrix S'.

[0032] Furthermore, the quality correction factor is:

[0033] If the value of the quality correction factor is smaller, it means that the user's rating is closer to the average rating, which means that the user's rating tends to be low or high overall, indicating that the user's rating is seriously affected by the quality of the scenic spot. The quality correction factor is used to modify the user-attraction rating matrix to cancel the influence of the scenic spot quality on the rating. The formula is: S u，i ″=S u，i ′(1+correction i ).

[0034] Furthermore, the implicit feedback data is used to modify the recommendation results. The specific method is as follows:

[0035] First, introduce the binary matrix L:

[0036]

[0037] It represents the target user’s preference for scenic spots, r is the user-scenic spot click matrix, r u,i is the number of clicks by user u on scenic spot i; therefore, the confidence level c is proposed u,i , which will increase as the number of clicks by user u on scenic spot i increases. α is a factor that controls the growth rate and is between (0,1):

[0038]

[0039] Among them, σ is the similarity between attractions, that is, the similarity between attractions that the user has clicked and attractions that have not been clicked:

[0040]

[0041] The preference result of the target user u can be obtained by performing matrix decomposition prediction on L. Taking the confidence level cu,i into consideration at the same time, the loss function is ∑ u，i c u，i (I u，i -X u Y i ) 2 +λ2(||I|| 2 +||Y|| 2 );

[0042] By minimizing the loss function, we can get X and Y, and the target user’s preference for each initially recommended attraction is:

[0043] L * u，i =X u Y i T

[0044] According to L *u,i Re-rank the recommendation results based on the size of .

[0045] In terms of privacy protection, considering that user information and attraction information are constantly updated, the federated learning method is implemented as follows:

[0046] Step A1: The user's implicit feedback data will be processed locally before uploading the results to the cloud. When the user generates a new rating, the cloud will download the scenic spot's characteristic vector view = (y1, y2, y3, y4), where y1, y2, y3, and y4 are the proportions of the scenic spot's cultural, consumption, internet celebrity, and environmental attributes. The user will modify its own user interest vector hobby based on the rating and view given locally, and only upload the updated gradient to the cloud.

[0047] The formula is:

[0048] hobby u =hobby u +β(c u,i (R u,i -L * u,i (hobby u T view i )))view i +λhobby u

[0049] The formula is to use the gradient descent method to combine user ratings with The difference is minimized to achieve the update of the user interest matrix, where β is the learning degree, c u,i is the confidence level between the user and the attraction, which is generally considered to be 1+αr u,i , α is [0,1], and λ is the regularization coefficient.

[0050] Step A2: After the user completes the local update, it uploads the update gradient to the cloud. At this time, the cloud should update the view of the rated scenic spot. The user only uploads the update gradient without uploading the specific data, which plays a role in privacy protection to a certain extent. The view update formula is as follows:

[0051] υiew i =υiew i +βf(u,i)+λυiew i

[0052] Among them, f(u,i) is the gradient of user u to scenic spot i

[0053] f(u,i)=(c u，i (R u，i -Lu，i *(hobby T υivew)))υiew i

[0054] Step A3: After both user data and attraction data are updated, recommendation results are regenerated and the user's interest vector is uploaded to the cloud. Because the user interest vector is a weighted vector, each party cannot directly receive the other party's original rating and corresponding private data. It cannot specifically express the user's preferences. Furthermore, a special quantitative standard is used, and the original data cannot be deciphered from the intermediate results. It can only reflect the user's general preferences, thus achieving a certain degree of privacy protection.

[0055] A personalized intelligent recommendation system for cultural and tourist attractions based on federated learning includes a user terminal u, servers T1, T2, and a main server P. The local terminal u is used to display recommendation results to users and is responsible for collecting users' explicit and implicit feedback data. When users generate new ratings, the local terminal is also responsible for updating the user's relevant data.

[0056] The server T1 requests user-related data from the user terminal to obtain a user cluster that is similar to the user in consideration of the user's natural characteristics and personalized selections, and sends the user cluster to the server T2 for use by T2 in making recommendations.

[0057] The server T2 is used to recommend the target user u after obtaining the similar user cluster given by T1. After obtaining the similar user cluster, it requests scores from each user in the cluster and requests implicit feedback data from the target user. The score and implicit feedback are combined to obtain an accurate recommendation result, and the recommendation result is sent back to the user terminal u.

[0058] The main server P is used to store all scenic spot data. It is responsible for transferring the required scenic spot data to server T2 and the user terminal after the user terminal u generates a score, and updating the rated scenic spots after the user terminal u completes the update and uploads the update gradient.

[0059] Furthermore, using the federated learning method, the scenic spot data is stored in the cloud and the user data is stored locally. When the user data is updated, only the update gradient is uploaded to the cloud without uploading the user's specific privacy data information. The cloud then uses the gradient information to update the global model and recommend personalized cultural and tourism scenic spot plans to users.

[0060] Furthermore, the user's implicit feedback data will be processed locally before the results are uploaded to the cloud. The scenic spot feature vectors of each scenic spot are combined into a scenic spot feature matrix and then uploaded to the cloud. When the user generates a new rating and the user-related information needs to be updated, the cloud splits the scenic spot matrix and sends the scenic spot feature vectors that user u has rated to user u locally to update the user interest vector hobby. At the same time, implicit feedback data and user ratings can both reflect the user's preferences. When updating the user interest vector, the implicit feedback data should be combined with the explicit feedback data. When user u has a high number of clicks on scenic spot i but gives a low rating, it is believed that the user is looking forward to this scenic spot i but is actually not satisfied with it. At this time, the user interest vector should be adjusted. Similarly, when user u has a low number of clicks on scenic spot i but gives a high rating, the user interest vector also needs to be significantly adjusted. Therefore, the update formula for the user interest vector is:

[0061] hobby u =hobby u +β(c u,i (R u,i -L * u,i (hobby u T view i )))view i +λhobby u

[0062] view is the characteristic vector of the scenic spot view = (y1, y2, y3, y4) T , y1, y2, y3, and y4 are the proportions of cultural, consumption, internet celebrity, and environmental attributes of the scenic spot, respectively;

[0063] The formula is to use the gradient descent method to combine user ratings with The difference is minimized to achieve the update of the user interest matrix, where β is the learning degree, c u,i is the confidence level between the user and the attraction, λ is the regularization coefficient, and the user interest vector is updated every time the user generates a rating;

[0064] When a user generates a new rating, the data of the rated attraction must also be updated. After uploading the user's updated gradient to the cloud, the updated attraction feature vector is:

[0065] υiew i =υiew i +βf(u,i)+λυiew i

[0066] Among them, f(u,i) is the gradient of user u to scenic spot i;

[0067] f(u,i)=(c u，i (R u，i -L u，i *(hobby T υivew)))υiew i

[0068] Only the gradient is uploaded instead of the specific data, thus protecting the user privacy.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] The present invention utilizes technologies such as collaborative filtering and federated learning, and takes into account methods such as user personalized needs to achieve accurate recommendations and privacy protection. It also combines implicit feedback data to alleviate the problem of system cold start. Compared with directly recommending to users, the present invention can not only achieve accurate recommendations that meet personalized needs, but also ensure the privacy and security of the recommendation process. It is suitable for large-scale promotion and can be applied to various travel recommendation platforms and social media. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the method of the present invention

[0072] Figure 2 Schematic diagram of a system in one embodiment of the present invention. DETAILED DESCRIPTION

[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0074] Please refer to Figure 1-2 The present invention delegates the recommendation process to a server T1, and the screening process to server T2. Through federated learning technology, user data is stored locally, and scenic spot data is stored in the main server P. This provides a personalized cultural and tourism scenic spot intelligent recommendation method based on federated learning.

[0075] like Figure 1As shown, user u first sends a recommendation request to server T1. Server T2 then requests a cluster of users similar to user u from server T1. Server T1 then requests user u's user interest vector hobby, the user's natural feature vector nature, and implicit feedback data. The user interest vector is a weighted vector. Neither party directly receives the other's original rating or corresponding private data. It cannot specifically express the user's preferences. Furthermore, using a special quantization standard, the original data cannot be deciphered from the intermediate results. It can only reflect the user's general preferences, thus achieving a certain degree of privacy protection. The server generates a recommendation result and returns it to user u. When user u gives a new rating, the user data must be updated locally. After the update is complete, the updated gradient is uploaded to the master server P, which then updates the rated attraction information.

[0076] In this embodiment, the user age preference vector age=(z1,z2,z3,z4) is defined as T ,z i ∈(0,1), and divide all users into two groups according to gender Where M represents the male group and F represents the female group, and the quantification rules are preset as follows:

[0077]

[0078] In this embodiment, server T1 generates a user cluster similar to user u, and the specific steps are as follows:

[0079] Step 1: Server T requests the user interest vector hobby=(x1,x2,x3,x4) and the user natural feature vector nature=(z1',z2',z3',z4',sex u ) T Among them, x1, x2, x3, and x4 are the user's preferences for cultural attractions, consumer attractions, Internet celebrity attractions, and environmental attractions, respectively. z1'=z1*1, z2'=z2*2, z3'=z3*3, z4'=z4*4. When user u is male, sex u = 0, when user u is female, sex u =1, z1, z2, z3, z4 are user age vector age = (z1, z2, z3, z4) TThe user ages are divided into z1, z2, z3, and z4, corresponding to the user's recognition of the labels youth, early adulthood, middle adulthood, and middle adulthood, respectively. Based on nature, we can calculate the natural feature similarity sim2 between different users, and based on hobby, we can calculate the interest similarity sim1 between different users. Taking sim1 and sim2 into account, we can obtain the user similarity sim.

[0080] Step 2: Calculate the interest similarity between different users. The user's interest vector is a weight vector. The weight vector does not have a large range of numerical deviations, and the values are already in the same dimension. Therefore, using the Pearson correlation coefficient, that is, pre-standardizing the user interest vector, is meaningless. Therefore, consider using the cosine similarity method to calculate the similarity of user interests:

[0081]

[0082] Step 3: Calculate the natural feature similarity sim2 between different users:

[0083] sim2 needs to reflect the similarity of different users in terms of age and gender. Euclidean distance can well reflect the differences between different users. The Euclidean distance formula between user u and user v is as follows:

[0084]

[0085] The larger the Euclidean distance, the greater the difference between user u and user v, and the smaller the natural feature similarity; the smaller the Euclidean distance, the smaller the difference between user u and user v, and the greater the natural similarity. Therefore, the formula for sim2 is as follows:

[0086]

[0087] Step A4: Get the similarity between different users and obtain similar user clusters:

[0088] The similarity formula between users is:

[0089] sim(u,v)=(1-α)sim1(u,v)+αsim2(u,v)

[0090] The top N users with the highest similarity to the target user are selected as similar user clusters and sent to server T2.

[0091] In this embodiment, server T2 generates recommendation results by performing the following steps:

[0092] Step 1: The n scenic spots rated by the user are formed into a user-scenic spot rating matrix S. For the matrix S, it is difficult for the target user to have rated the scenic spots rated by other users, which leads to vacancies in the target user's rating. The target user's rating is predicted by the matrix decomposition method. It is expected that the matrix S can be decomposed into: S = p × q. After the matrix decomposition, the eigenvector of user u mapped to the k-dimensional space is p u , the characteristic vector of scenic spot n is q i , the predicted score is: Let the loss function When p and q are minimized, BP neural network is used for training. The key is how to obtain the learning rate at different time stages. By combining Gaussian kernel method with time similarity, the time set T of the first n-1 iterations is obtained. n-1}, get the learning rate of the nth iteration:

[0093]

[0094] Multiplying p and q together can obtain a denser user-attraction rating matrix S'.

[0095] Step 2: Considering the impact of scenic spot quality on scenic spot ratings, introduce a quality correction factor:

[0096]

[0097] If the value of the quality correction factor is smaller, it means that the user's rating is closer to the average rating, which means that the user's rating tends to be low or high overall, indicating that the user's rating is seriously affected by the quality of the scenic spot.

[0098] The quality correction factor is used to modify the user-attraction rating matrix to cancel the influence of attraction quality on the rating. The formula is: S u，i ″=S u，i ″(1+correction i )

[0099] Step 3: Observe the row in S" that belongs to the target user, and select the top N attractions to recommend to the user based on the ratings. The order of recommendations is determined by the user's implicit feedback data.

[0100] First, we introduce the binary matrix L, which represents the target user’s preference for attractions, r is the user-attraction click matrix, r u,i is the number of clicks by user u on scenic spot i. If user u has not clicked on scenic spot i, then L u,i =0, if user u has clicked on scenic spot i, L u,i=1, we believe that user u is likely to like scenic spot i. Of course, the increase in the number of clicks means that user u likes scenic spot i more obviously, so the confidence level of user liking scenic spot i is proposed:

[0101]

[0102] Among them, σ is the similarity between unclicked attractions and clicked attractions:

[0103]

[0104] By performing matrix decomposition prediction on L while considering the confidence level, we can get a more accurate user preference. We expect to decompose L into L = X × Y, and the loss function is:

[0105] Σ u，i c u，i (I u，i -X u Y i ) 2 +λ2(||I|| 2 +||Y|| 2 )

[0106] By minimizing the loss function, we can get X, Y. The target user’s preference for each initially recommended attraction is L. * u，i =X u Y i T , re-rank the initially recommended attractions according to the results of L*u,i.

[0107] In this embodiment, the update of the user terminal u and the main server P is performed according to the following steps:

[0108] Step 1: When user u generates a new rating, the main server P sends the rated attraction information view = (y1, y2, y3, y4) to user u locally, where y1, y2, y3, and y4 are the proportions of the attraction's cultural, consumer, influencer, and environmental attributes, respectively. The user interest vector of user u is updated locally. Both implicit feedback data and user ratings reflect user preferences. When updating the user interest vector, implicit feedback data should be combined with explicit feedback data. The user interest vector is updated using the following formula:

[0109] hobby u =hobby u +β(c u,i (R u,i -L * u,i (hobby uT view i )))view i +λhobby u

[0110] Among them, hobby is the user interest vector, and the formula is to combine the user rating and The difference is minimized to achieve the update of the user interest matrix, where β is the learning degree, c u,i is the confidence level between the user and the attraction, which is generally considered to be 1+αr u,i , α is [0,1], and λ is the regularization coefficient.

[0111] Step 2: After the user-side update is completed, the updated gradient is uploaded to the main server P, and the scenic spot data view of the rated scenic spot is updated at P. The update formula is as follows:

[0112] view i =view i +βf(u,i)+λview i

[0113] Among them, f(u,i) is the gradient of user u to scenic spot i

[0114] f(u,i)=(c u，i (R u，i -L u，i * (hobby T υivew)))υiew i

[0115] Uploading only the gradient without uploading the specific data also protects user privacy to a certain extent.

[0116] The present invention also provides a recommendation system for implementing the above method, such as Figure 2 As shown, it includes user terminal u, main server P, server T1, and server T2.

[0117] The local end u is used to display the recommendation results to the user and is responsible for collecting the user's explicit and implicit feedback data. When the user generates a new rating, the local end is also responsible for updating the user's relevant data.

[0118] The server T1 requests user-related data from the user terminal to obtain a user cluster that is similar to the user in consideration of the user's natural characteristics and personalized selections, and sends the user cluster to the server T2 for use by T2 in making recommendations.

[0119] The server T2 is used to recommend the target user u after obtaining the similar user cluster given by T1. After obtaining the similar user cluster, it requests scores from each user in the cluster and requests implicit feedback data from the target user. The score and implicit feedback are combined to obtain an accurate recommendation result, and the recommendation result is sent back to the user terminal u.

[0120] The main server P is used to store all scenic spot data. It is responsible for transferring the required scenic spot data to server T2 and the user terminal after the user terminal u generates a score, and updating the rated scenic spots after the user terminal u completes the update and uploads the update gradient.

[0121] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A personalized intelligent recommendation method for cultural and tourist attractions based on federated learning, characterized by: The following steps are involved: Define the user's natural feature vector nature and interest feature vector hobby, then consider the similarity sim between users by combining the natural features and interest features to obtain a user cluster similar to the target user u; Integrate the attraction ratings given by all users in the user cluster similar to the target user u into the user-attraction rating matrix S; Use quality correction factor correction i First, S is modified, and then matrix decomposition prediction is performed to obtain the top n desired attractions for the target user u. Then, combined with implicit feedback data, the order of the recommendation results is modified to obtain the final recommendation result. The user-attraction rating matrix S is corrected and matrix decomposition is predicted. The specific method is as follows: For the user-attraction rating matrix, it is expected to decompose the matrix S into: S m×n =p m×k ×q k×n ; Assume that user u’s rating of scenic spot i is S u,i , after matrix decomposition, the feature vector of user u mapped to the k-dimensional space is p u , the characteristic vector of scenic spot n is q i , the predicted score is: Introducing the regularization factor λ, the loss function is: Use BP neural network to train and minimize the loss function. Let the time set of the first n-1 iterations be T = {t0, t1, ..., t n-1 }, then the learning rate γ for the nth iteration i (t i )for: When the loss function reaches its minimum, we get p and q. Multiplying p and q together gives us the dense user-attraction rating matrix S'. The mass correction factor: The quality correction factor is used to modify the user-attraction rating matrix to cancel the influence of attraction quality on the rating. The formula is: S u，i ″=S u，i ′(1+correction i ); The implicit feedback data is used to modify the recommendation results. The specific method is as follows: First, introduce the binary matrix L: It represents the target user’s preference for scenic spots, r is the user-scenic spot click matrix, r u，i is the number of clicks by user u on scenic spot i; therefore, the confidence level c is proposed u，i , which will increase as the number of clicks by user u on scenic spot i increases. α is a factor that controls the growth rate and is between (0, 1): Among them, σ is the similarity between attractions, that is, the similarity between attractions that the user has clicked and attractions that have not been clicked: Perform matrix decomposition prediction on L to obtain the preference result of target user u. Consider the confidence level cu,i at the same time, and obtain the loss function as follows: ∑ u，i c u，i (I u，i -X u Y i ) 2 +λ2(||I|| 2 +||Y|| 2 ); Minimize the loss function and get X and Y, then the target user's preference for each initially recommended attraction L * u，i =X u T Y i According to L * u，i Re-rank the recommendation results based on the size of .

2. The personalized intelligent recommendation method for cultural and tourist attractions based on federated learning according to claim 1 is characterized in that: The definition of the user's natural feature vector nature and interest feature vector hobby is as follows: Define user age preference vector age=(z1,z2,z3,z4) T ,z i ∈(0,1), and divide all users into two groups according to gender M=(u1,u2,u3,u4,...u n ,), F=(u1,u2,u3,u4,...u n ,), where M represents the male group, F represents the female group, and the quantization rules are preset; The user's nature vector is calculated based on the quantization rule, where z1'=z1*1,z2'=z2*2,z3'=z3*3,z4'=z4*4. When user u is male, sex u = 0, when user u is female, sex u =1; Divide attractions into four categories: cultural, consumer, internet celebrity, and environmental; and construct the user interest vector hobby = (x1, x2, x3, x4) T ,x i ∈(0,1), where x1, x2, x3, and x4 are the user’s preferences for cultural attractions, consumer attractions, internet celebrity attractions, and environmental attractions, respectively; the sum of x1, x2, x3, and x4 is 1.

3. The personalized intelligent recommendation method for cultural and tourist attractions based on federated learning according to claim 1 is characterized in that: The similarity sim between users is considered by combining natural features and interest features to obtain a user cluster similar to the target user u, specifically as follows: Step A1: The inter-user similarity sim is obtained by the user interest similarity sim1 and the user natural similarity sim2. First calculate sim1: It is meaningless to use the Pearson correlation coefficient to standardize the user interest vector in advance, so consider using the cosine similarity method to calculate the similarity of user interests: Step A2: Calculate sim2 again, based on the Euclidean distance to reflect the differences between different users. The Euclidean distance formula between user u and user v is as follows: The larger the Euclidean distance, the greater the difference between user u and user v, and the smaller the similarity; the smaller the Euclidean distance, the smaller the difference between user u and user v, and the greater the similarity; therefore, the formula for sim2 is as follows: Finally get sim: sim(u,v)=(1-α)sim1(u,v)+αsim2(u,v).

4. A personalized intelligent recommendation system for cultural and tourist attractions based on federated learning, characterized by: The system is implemented using the personalized intelligent recommendation method for cultural and tourism attractions based on federated learning as described in any one of claims 1 to 3, including a user terminal u, a server T1, a server T2, and a main server P; the user terminal u is used to display recommendation results to users and is responsible for collecting user explicit and implicit feedback data. When a user generates a new rating, the user terminal u is also responsible for updating the user's relevant data; The server T1 requests user-related data from the user terminal to derive a user cluster that is similar to the user, taking into account the user's natural characteristics and personalized choices, and sends the user cluster to the server T2 for use by T2 in making recommendations; The server T2 is used to recommend the target user u after obtaining the similar user cluster provided by T1. After obtaining the similar user cluster, it requests scores from each user in the cluster and requests implicit feedback data from the target user. It combines the scores with the implicit feedback to obtain an accurate recommendation result and sends the recommendation result back to the user terminal u. The main server P is used to store all scenic spot data; it is responsible for transferring the required scenic spot data to the server T2 and the user terminal after the user terminal u generates a score, and updating the scored scenic spots after the user terminal u completes the update and uploads the update gradient.

5. The personalized cultural and tourism attraction intelligent recommendation system based on federated learning according to claim 4 is characterized in that: Using the federated learning method, the scenic spot data is stored in the cloud and the user data is stored locally. When the user data is updated, only the update gradient is uploaded to the cloud without uploading the user's specific privacy data information. The cloud then uses the gradient information to update the global model and recommend personalized cultural and tourism scenic spot plans to users.

6. The personalized cultural and tourism attraction intelligent recommendation system based on federated learning according to claim 5 is characterized in that: The user's implicit feedback data will be processed locally before the results are uploaded to the cloud. The scenic spot feature vectors of each scenic spot are combined into a scenic spot feature matrix and then uploaded to the cloud. When the user generates a new rating and the user-related information needs to be updated, the cloud splits the scenic spot matrix and sends the scenic spot feature vectors that user u has rated to user u locally to update the user interest vector hobby. At the same time, both implicit feedback data and user ratings reflect the user's preferences. When updating the user interest vector, the implicit feedback data should be combined with the explicit feedback data. When user u has a high number of clicks on scenic spot i but gives a low rating, it is believed that the user is looking forward to this scenic spot i but is actually not satisfied with it. At this time, the user interest vector should be adjusted. Similarly, when user u has a low number of clicks on scenic spot i but gives a high rating, the user interest vector also needs to be significantly adjusted. Therefore, the update formula for the user interest vector is: hobby u hobby u +β(c u，i (R u，i -L * u，i (hobby u T view i )))view i +λhobby u view is the characteristic vector of the scenic spot view = (y1, y2, y3, y4), where y1, y2, y3, and y4 are the proportions of the cultural, consumption, internet celebrity, and environmental attributes of the scenic spot respectively; The formula is to use the gradient descent method to combine user ratings with The difference is minimized to achieve the update of the user interest matrix, where β is the learning degree, c u，i is the confidence level between the user and the attraction, λ is the regularization coefficient, and the user interest vector is updated every time the user generates a rating; When a user generates a new rating, the data of the rated attraction must also be updated. After uploading the user's updated gradient to the cloud, the updated attraction feature vector is: view i =view i +βf(u,i)+λview i Among them, f(u, i) is the gradient of user u to scenic spot i; f(u,i)=(c u，i (R u，i -L u，i * (hobby T view)))view i Only the gradient is uploaded instead of the specific data, thus protecting the user privacy.

Images