A MEC cache optimization method based on streaming media

Abstract

The present invention belongs to the field of communication technology, and specifically relates to a method for optimizing MEC cache based on streaming media, including: obtaining streaming media preference data through an MEC server, predicting users based on the streaming media preference data, filling target users based on the predicted results, using a hierarchical analysis method to calculate the user's scoring weight, and finally using a particle swarm algorithm to determine the list of files that need to be stored in the MEC server. Through the method of the present invention, a faster network service that better meets user needs can be provided. MEC research includes network planning, resource allocation, interference coordination, and cache planning. The present invention combines the characteristics of "streaming media" and further optimizes the cache planning according to the different preferences of different users for streaming media content. The present invention formulates a cache optimization scheme so that the cached files of the MEC server can bring the best experience to surrounding users under limited cache space.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile communications, and in particular relates to a MEC cache optimization method based on streaming media. Background Art

[0002] With the continuous development of mobile networks, people's demand for networks is also increasing. According to statistics, half of mobile data is streaming data. With the rise of major self-media platforms in recent years, people's demand for network services has continued to increase. LTE faces considerable challenges in terms of data tolerance and data transmission speed. MEC was proposed by the European Telecommunications Standards Institute (ETSI) in 2014. It moves computing and caching functions from the network center to the network edge, effectively reducing file transmission delays, alleviating the burden on the core network, and providing a more efficient network service.

[0003] In MEC edge caching, active caching is a method that takes advantage of this traffic dynamics by actively caching popular content during off-peak hours, thereby reducing the demand for peak traffic. Since the cache units of edge caching are deployed at the edge of the network, more information can be used to improve cache efficiency. Currently, when MEC servers actively cache, there is no good caching solution for content that needs to be cached in advance, and the hit rate of target files is low. During peak traffic periods, since the cached content has no practical significance for users in the area where the MEC server belongs, it causes a waste of resources and slow file transfer speeds. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a MEC cache optimization method based on streaming media to solve the problems of low hit rate, slow transmission speed and waste of resources in MEC cache space in the current MEC cache optimization scheme.

[0005] The present invention provides a MEC cache optimization method based on streaming media, which is achieved by the following specific technical means:

[0006] First, the MEC cache optimization platform collects the streaming media preference data obtained by the MEC server. The streaming media preference data includes: user attributes, number of users, projects participated by users, project attributes, user ratings of projects, the size of the storage space occupied by projects, the frequency of users accessing streaming media, and other information;

[0007] The project-attribute matrix and the project-user rating matrix are constructed by using the obtained streaming media preference data, and the project-user rating matrix and the project-attribute matrix are fused to obtain the project-user attribute preference matrix;

[0008] The comprehensive similarity of users is calculated based on the item-user rating matrix and the item-user attribute preference matrix:

[0009] The project-user attribute preference similarity is calculated using the cosine similarity formula based on the project-user attribute preference matrix; the project-user rating similarity, user preference similarity and confidence are calculated based on the project-user rating matrix; the user-rating similarity is calculated based on the project-user rating similarity, confidence and user preference similarity; the user comprehensive similarity is calculated based on the project-user attribute preference matrix and the project-user attribute preference similarity;

[0010] If there are items in the item-user rating matrix that the user has not participated in rating, the K nearest neighbor users of the user are determined based on the comprehensive similarity, and the user's rating of the item is predicted based on the ratings of the K nearest neighbor users, and the predicted rating is filled into the item-user rating matrix;

[0011] According to the frequency of users' access to streaming media data, users are divided into multiple levels, and the scoring weights of users at different levels are determined using the analytic hierarchy process (AHP).

[0012] According to the user's rating weight and the filled project-user rating matrix, the ratings of all users for each project are calculated, and the final score of each project is determined. The projects are arranged into a file list according to the final score;

[0013] The particle swarm algorithm is used to calculate the best cache solution based on the order of the items in the file list, the cache space of the MEC server, the size of the storage space occupied by the items, and the final score of the items;

[0014] According to the optimal caching scheme, the MEC server caches the streaming data in advance.

[0015] The present invention has at least the following beneficial effects:

[0016] By integrating project-user attributes and project-user scores, reference factors are added to make the prediction results more accurate. By calculating the comprehensive similarity of users, the nearest neighbor users who are most similar to the target customers are found, which improves the accuracy of the prediction of the target user score. According to the comprehensive similarity of users combined with the project-user score matrix, the scores of the projects in which the target user has not participated are predicted and filled, which improves the accuracy of the project-user score matrix and provides more complete data for formulating cache optimization schemes. The particle swarm algorithm perfectly combines the characteristics of the streaming media preference data itself and the cache space of the MEC server, making the calculated cache optimization scheme the most reasonable. The MEC cache optimization scheme implemented by the present invention greatly improves the hit rate of streaming media data, while increasing the transmission speed and reducing the resource waste of the MEC server cache space. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1A flowchart of an embodiment of the present invention;

[0018] Figure 2 A comparison chart of file transmission time between the present invention and the comparative solution in different scenarios;

[0019] Figure 3 A comparison diagram of the movie cache miss probability of the present invention and the comparative solution in different scenarios;

[0020] Figure 4 This is the influence diagram of the project-user attribute preference similarity trade-off factor α on MAE of the present invention. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present invention will be described clearly and in detail below in conjunction with the drawings of the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention and cannot be used to limit the scope of the present invention.

[0022] See also Figure 1 ,First, according to the characteristics of the MEC server, the ,streaming media preference data obtained by the MEC server is collected through the ,MEC cache optimization platform, and according to the characteristics of the ,streaming media preference data, it is processed into the item-user rating matrix and the ,item-attribute matrix:

[0023] Through the streaming media preference data, we can get B = {B 1 ,B 2 ...B t}, B is the attribute set of the project, i = {i 1 ,i 2 ...i n} is a set of items, u={u 1 ,u 2 ...u m} is the user set. As shown in Table 1, we get an n*t-dimensional item-attribute matrix C n*t , n represents the number of items, and t represents the number of attributes in the items. As shown in Table 2, we can obtain an n*m-dimensional item-user rating matrix C n*m , where m represents the number of users, n represents the number of projects, and C ka Indicates the user's rating of the project, and the user uses u m Indicates that the project is represented by i n Indicates that the attribute is represented by B t express:

[0024] Table 1 Project-attribute matrix C n*t

[0025]

[0026]

[0027] Table 2 Project-user rating matrix C n*m

[0028]

[0029] If the project i k Has attribute B S (B S ∈B), then C ks =1(s∈t, k∈n), if the item i k No attribute B S , C ks =0, if user u m For this project k Rating, then C ka Equal to the rating, if user u m No items k Rating, then C ka Equal to 0, the above m, n, t, C ka and C ks The value of is determined according to the streaming preference data obtained by the MEC server.

[0030] The project-user rating matrix and the project-attribute matrix are combined to obtain the project-user attribute preference matrix. The project-user attribute preference similarity is calculated using the cosine similarity formula based on the project-user attribute preference matrix:

[0031] According to formula (1), the project-user rating matrix and the project-attribute matrix are fused to obtain the project-user attribute preference matrix, as shown in formula (1):

[0032]

[0033] p as Represents user u a For Project I k Medium Attribute B s The degree of preference, p as ∈[0,1],r ak Represents user u a For Project I k Rating, I s Represents user u a Among all the items involved in the scoring, those with attribute B s The project set, B represents the attribute set of the project, and the project-user attribute preference matrix P is obtained by formula (1): m*t , B s ∈B represents the attribute B in the attribute set B of the item s , t represents the number of attributes in the project, as shown in Table 3:

[0034] Table 3 Project-user attribute preference matrix P m*t

[0035]

[0036] Among them, Ai ks Indicates project i k Medium Attribute B s , p as Represents user u a For Project I k Medium Attribute B s degree of preference;

[0037] The comprehensive similarity of users is calculated based on the item-user rating matrix and the item-user attribute preference matrix:

[0038] 1. Calculate the item-user attribute preference similarity using the cosine similarity formula based on the item-user attribute preference matrix:

[0039]

[0040] In formula (2), p as Represents user a's response to item i k Medium Attribute B s The degree of preference, p us Represents user u's response to item i k Medium Attribute B S The degree of preference, sim ak Represents the relationship between user a and user u for item i k Medium Attribute B s The preference similarity of k represents a collection of items, B s ∈B represents the attribute B in the attribute set B of the item s , t represents the number of attributes in the project.

[0041] 2. Calculate the item-user rating similarity, user preference similarity and confidence based on the item-user rating matrix:

[0042] 1) Item-User Rating Similarity

[0043] The greater the difference between the ratings of different users for the same content, the lower their similarity. The sigmoid function is used to represent the similarity of ratings between different users:

[0044]

[0045] In formula (3), sim 1(u,v,i) represents the similarity of the ratings of users u and v on the projects they jointly participated in. u represents a user, v represents another user, and P vi represents the rating of user v on item i, P ui I represents the score of item i by user u, and the set of items that user v participates in scoring is represented by I v Indicates that the set of items that user u participates in rating is represented by I u Indicates that I u ∩I v represents the common items that users u and v participate in rating, where i∈I u ∩I v , i represents a project in the set of common projects participated by users u and v, and i here only represents the meaning in formula (3).

[0046] 2) User preference similarity

[0047] According to the rating habits of different users, the following calculation formula is introduced to measure the similarity of users' interest preferences for the same project:

[0048]

[0049] In formula (4), sim 2 (u,v,i) represents the similarity of rating preferences of users u and v for the projects they jointly participated in, where u represents a user and v represents another user. ui represents the rating of item i by user u, P vi represents the rating of user v on item i, Represents the average ratings of users u and v on all participating projects, where i∈I u ∩I v , i represents a project in the set of common projects participated by users u and v. Here, i only represents the meaning in formula (4).

[0050] 3) Confidence

[0051] There are few items that are rated in common among different users. Although the calculated rating similarity is very high, they are not necessarily similar. Here, the user similarity is corrected by introducing the Jaccard coefficient. The Jaccard coefficient is shown in formula (5):

[0052]

[0053] In formula (5), sim 3 (u,v) represents the confidence between user u and user v, I u ∩I v represents the common items that users u and v participate in rating, I u∪I v Represents all items rated by user u or user v.

[0054] 3. Calculate the user-rating similarity based on the item-user rating similarity, confidence, and user preference similarity:

[0055]

[0056] in:

[0057]

[0058] sim u (u,v) represents the rating similarity between user u and user v, sim represents the fusion similarity of item-user rating similarity, user preference similarity and confidence, where C uv =I u ∩I v , represents the set of projects that users u and v jointly participate in rating, i represents a project in the set of common projects participated by users u and v, and i here only represents the above meaning in formula (7).

[0059] 4. Calculate the user comprehensive similarity based on the user-rating similarity and item-user attribute preference similarity:

[0060] Linearly weight the item-user rating similarity and item-user attribute similarity obtained in 2-3 above to obtain a user comprehensive similarity sim':

[0061] sim'=(1-α)sim u +αsim ak (8)

[0062] sim u Represents the item-user rating similarity between two users, sim ak represents the item-user attribute preference similarity between two users, and α represents the proportion of item-user attribute preference similarity. The value of α is shown in formula (9):

[0063]

[0064] See also Figure 4 ,The impact of item-user attribute similarity trade-off factor α on MAE,In order to weigh the weights of the two similarities, a trade-off factor α is defined. α is between [0,1]. Different values ​​of α have different recommendation effects. Different values ​​of α in [0,1] are taken to analyze the impact of α on the recommendation results. According to Figure 4It can be found that the value of α can indeed affect the recommendation effect. When 0.3<α<0.5, the change tends to be gentle, indicating that the recommendation effect is better in this range. Because the present invention integrates project-user attributes and user-project ratings, if the value of parameter α is too small, the result will be biased towards the project-user rating similarity, which cannot well reflect the impact of project-user attributes on the recommendation effect. If the value of α is too large, the result will be too biased towards the project-user attribute similarity. Through experiments, it can be found that the recommendation effect is best when α is 0.4.

[0065] According to the comprehensive similarity between different users, one of the users is selected as the target user in turn to determine whether the user has any project that he has not participated in rating. If the user has any project that he has not participated in rating, the remaining users and the target user are sorted from large to small according to the comprehensive similarity. Whether the user participates in rating the project is determined according to the corresponding C in the project-user rating matrix in Table 2. ka Is it 0? If C ka is 0, then user u a Not involved in project 0 The first K users who are most similar to the target user are taken as neighboring users. According to the neighboring users' evaluation of the target user not participating in project i in the project-user rating matrix, 0 The score of the target user's non-participated project i 0 Perform rating prediction, fill the predicted results into the project-user rating matrix, and complete the filling of the project-user rating matrix. The K nearest neighbor users can be randomly selected by technicians in this field through experiments. The calculation of rating prediction is shown in formula (10):

[0066]

[0067] In formula (10), Indicates the target user Not participating in the scoring project 0 Rating prediction of represents the average rating of user u on all projects he has participated in, U' represents the set of K users most similar to the target user u, Represents the neighbor user u''s response to item i 0 Rating, represents the average rating of all participating projects by neighboring user u', sim'(u,u') represents the comprehensive similarity between target user u and neighboring user u', and the predicted result Fill in the target user u's rating matrix for item i 0 In terms of ratings, i 0represents the project that target user u has not participated in rating, where K can be set according to the experimental results. In the embodiment of the present invention, K is set to 50. For example, when the project-user rating matrix is ​​not filled, the target user u has no participation in rating the project i. 0 The score of item i is 0. After filling the item matrix, the target user u has a positive opinion on item i. 0 The rating is

[0068] By integrating project-user attributes and project-user ratings, reference factors are added to make the prediction results more accurate. By calculating the comprehensive similarity of users, the nearest neighbor users who are most similar to the target user are found, which improves the accuracy of the target user rating prediction. Based on the comprehensive similarity of users combined with the project-user rating matrix, the ratings of the projects in which the target user has not participated are predicted and filled, which improves the accuracy of the project-user rating matrix and provides more complete data for formulating cache optimization solutions.

[0069] According to the frequency of users' access to streaming media data, users are divided into multiple levels. In the embodiment of the present invention, users are divided into 9 levels. The analytic hierarchy process (AHP) is used to determine the scoring weights of users at different levels. According to the scoring weights combined with the filled item-user scoring matrix, the scores of all users on each item are averaged to obtain the final score of each item. The items are arranged into a file list from high to low according to the final score. For example, when the level of user a is 1 and the level of user b is 3, user b is slightly more important than user a. According to the pairwise comparison matrix scaling method, it can be concluded that user a is more important than user b. 13 For 3:

[0070] 1. According to the importance of user a of level z relative to user b of level j, users of different levels are compared pairwise, and a pairwise comparison matrix is ​​constructed as shown in (11):

[0071]

[0072] In the above matrix, a zj It represents the importance of user a of level z relative to user b of level j, and has the following properties:

[0073]

[0074] The value of a is selected according to the numerical rules in Table 4:

[0075] Table 4 Pairwise comparison matrix scaling method

[0076]

[0077] 2. Normalize the comparison matrix in Section 1 to obtain the column-normalized pairwise comparison decision matrix:

[0078]

[0079] In formula (13), n represents how many levels all users are divided into. For the convenience of description, n is 9 in this embodiment.

[0080] 3. Add the matrix columns processed in Section 2:

[0081]

[0082] In formula (14), n represents the number of levels that all users are divided into. In this embodiment, n is 9. A feature vector is calculated by (14)

[0083] 4. Normalize the vector

[0084]

[0085] In formula (15), n represents the number of levels that all users are divided into. In this embodiment, n is 9. z Represents the user rating weight of the zth level.

[0086] 5. Matrix consistency check

[0087] The consistency check steps are as follows:

[0088] A: Calculate the maximum eigenvalue λ of the pairwise comparison matrix max .

[0089] B: Find the consistency index:

[0090]

[0091] When CI=0, it means that the constructed pairwise comparison matrix is ​​consistent. The smaller the CI, the better the consistency. In formula (16), n represents how many levels all users are divided into. In this section of this embodiment, n is 9.

[0092] C: The corresponding random consistency index RI is obtained by taking the average through random simulation. RI is the standard consistency protocol and is obtained by looking up the table.

[0093] D: Calculate the consistency ratio:

[0094]

[0095] When CR<=0.1, determine W zis the user's rating weight. When CR>0.1, it is reconstructed into paired comparison evidence, and this section 2-5 is repeated until CR<=0.1. The values ​​of z and j in this embodiment are integers from 1 to 9.

[0096] According to the weight value W obtained above z ,The user’s scoring weight is combined with the filled item-user scoring matrix, the scores of all users on each item are averaged, and the final score of each item is obtained. The items are arranged into a file list from high to low according to the final score. The user’s scoring weight is calculated by the hierarchical analysis method, and the influence of the user’s access frequency on the comprehensive score of the item is considered to improve the accuracy of the data. In the cache scheme, the hit rate of the cached content is improved. For example, for existing users q, l, e, the scoring weight of users q, l, e is determined according to their levels. W zq , W zl , W ze are the rating weights of users q, l, and e, respectively, and C ka For user a on item i k , then users q, l, e have a positive opinion on item i 1 The score is expressed as C 1q , C 1l , C 1e , users q, l, e have different opinions on item i 2 The score is expressed as C 2q , C 2l , C 2e , users q, l, e have different opinions on item i 3 The score is expressed as C 3q , C 3l , C 3e , Project i 1 The final score is Project i 2 The final score is Project i 3 The final score is According to project i 1 、i 2 、i 3 The final scores are ranked from high to low, and a file list is generated. The file list includes the project, the size of the storage space occupied by the project, and the final score of the project. The size of the storage space occupied by the project is obtained through the MEC server. Here is just a brief description of how to obtain the final score of the project, and it cannot be used to limit the scope of protection of the present invention.

[0097] The particle swarm algorithm is used to calculate the optimal cache solution by combining the order of the items in the file list, the cache space of the MEC server, the size of the storage space occupied by the items, and the final score of the items:

[0098] A1: According to the number of items in the file list, take the corresponding particle swarm number N, the corresponding space dimension d, and the corresponding number of iterations k. The space dimension d is equal to the number of items n in the file list. For the convenience of expression, in this section of the particle swarm algorithm, n represents the number of items in the file list. The speed and position of the particles are randomly initialized by the rand() function in MATLAB. Each dimensional space velocity vector of each particle is randomly initialized to a real number, and each dimensional space position vector of each particle is randomly initialized to 0 or 1. In this embodiment, the number of items is 10, the number of particle swarms is 100, the number of space dimensions is 10, and the number of iterations is 100. Those skilled in the art can appropriately change the values ​​according to the actual effect;

[0099] A2: According to the cache space of the MEC server, the size of the storage space occupied by the project and the final score of the project, the objective function of the particle swarm and the limiting function of the objective function are constructed:

[0100]

[0101]

[0102] In formula (18), n represents the total number of items in the file list, and maxf(X 1 ,X 2 ,...,X N ) represents the total score of all items in the cache solution. In the particle swarm algorithm, maxf(X 1 ,X 2 ,...,X N ) represents the fitness value of the particle. In the objective function of the particle swarm, maxf(X 1 ,X 2 ,...,X N ), X N represents the set of position increments of particle N in d-dimensional space, x i A 0-1 list representing whether to store or not store the i-th item in the file list, x i With X N correspond, As the objective function maxf(X 1 ,X 2 ,...,X N ), w i Indicates the size of the storage space occupied by the project, V indicates the size of the MEC server cache space, C i represents the final score of item i, when caching the i-th item in the file list x i Equal to 1, when the i-th item in the file list is not cached x iis equal to 0, i represents the order of the items in the file list. For example, when i=1, it represents the item with the highest final score in the file list. When i=2, it corresponds to the item with the second highest final score in the file list. For example, X 1 represents the position increment of the first particle in the d-dimensional space. Since each dimensional space position vector of each particle is randomly initialized to 0 or 1 in step A1, the position increment X of the first particle at the beginning of the first iteration is 1 The initialization position is the same, if X 1 It is represented as [0 0 1 1 0], then X 1 It means that the position increment of particle 1 in the first dimension of the 5-dimensional space is 0, the position increment in the second dimension is 0, the position increment in the third dimension is 1, the position increment in the fourth dimension is 1, and the position increment in the fifth dimension is 0. Since the spatial dimension d is equal to the number of items in the file list, X 1 The corresponding x i Represented as {x 1 =0,x 2 =0,x 3 =1,x 4 =1,x 5 =0}, then the corresponding cache scheme is that the MEC server does not cache the first item, does not cache the second item, caches the third item, caches the fourth item, and does not cache the fifth item;

[0103] A3: Calculate the particle position increment:

[0104]

[0105] in, X N represents the set of position increments of particle N in d-dimensional space, is the position increment of particle N in d-dimensional space after the kth iteration, represents the velocity of particle N in d-dimensional space, is a uniform random number between (-1, +1), For is a variable, a hyperbolic tangent function with a value between (-1, +1), the speed of the first iteration is the initial speed, X N A 0-1 list is obtained through formula (19);

[0106] A4: According to the objective function and the particle's position increment X in d-dimensional space N The fitness value of each particle is calculated, and the highest fitness value of the particle is taken as the fitness value of the group particles;

[0107] A5: Determine whether the fitness value of the particle satisfies the limiting function. When the fitness value of the particle does not satisfy the limiting function, set the fitness value of the particle to 1. When the fitness value of the particle satisfies the limiting function and the fitness value of the particle is greater than its historical best fitness value, replace the historical best fitness value of the particle with the current fitness value of the particle. When the historical best fitness value of a particle is greater than the historical best fitness value of the group particles, replace the historical best fitness value of the group particles with the historical best fitness value of the particle and record the position increment X corresponding to the historical best fitness value of the current group particles. N As a feasible solution;

[0108] A6: Use formula (20) to update the particle's velocity, and formula (21) to update the particle's position:

[0109]

[0110]

[0111] in, represents the velocity of particle N in the d-dimensional space at the k+1th iteration, w represents the inertia factor, which in this embodiment is a random number between [0, 1], specifically 0.5, t1 and t2 refer to the learning factors, c1=c2=2, r1 and r2 are random numbers between (0, 1), Indicates the position of the historical best fitness value of particle N in the d-dimensional space. represents the position of particle N in d-dimensional space, It indicates the position of the historical best fitness value of the group particles in the d-dimensional space after the kth iteration. is the position of particle N in the d-dimensional space after the k+1th iteration.

[0112] Repeat the above steps A3-A6 until the iteration ends, and finally output the position increment X corresponding to the historical best fitness value of the group particles N , because the spatial dimension defined is equal to the number of items in the file list and the spatial dimension corresponds to the item number in the file list, so X N Equivalent to x i , according to X N Determine the corresponding cache list and obtain the best cache solution. According to the best cache solution, the MEC server caches the streaming media data in advance. N The final output is a list of 0 or 1, which is the same as the list x of cached or non-cached items in the file list. i Corresponding equal.

[0113] The designed particle swarm algorithm perfectly combines the characteristics of the streaming preference data itself and the cache space of the MEC server, and at the same time combines with the obtained file list to make the cache solution more reasonable.

[0114] See also Figure 2 , Figure 2 The following is a comparison chart of file transfer time in different scenarios. As the number of accessed files increases, the optimization effect will be more obvious. When the number of accessed files is large, a considerable amount of transfer time can be saved. The optimized cache solution has more obvious advantages than the previous single cache solution, which will save more transfer time and bring higher efficiency. The more files there are, the more obvious the difference in comparison effect.

[0115] See also Figure 3 , Figure 3 This is a comparison chart of the probability of file cache misses in different scenarios. Although there are certain fluctuations in the experimental data, the overall trend is stable. The cache solution provided by the present invention reduces the miss rate of streaming media data. The MEC cache optimization solution proposed by the present invention improves the traditional cache solution and can provide users with more efficient services. Due to the excessive number of streaming media data items, this embodiment only lists the movie items in the streaming media data.

[0116] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the present invention. Without departing from the spirit and essence of the method of the present invention, those skilled in the art can make various corresponding changes according to the method of the present invention, which all fall within the scope of protection of the claims of the method of the present invention.

Claims

1. A method for optimizing MEC cache based on streaming media, characterized in that: include: S1: The MEC cache optimization platform collects streaming media preference data obtained by the MEC server, and the streaming media preference data includes: the number of users, the attributes of users, the number of projects and the user's rating of the project, the size of the storage space occupied by the project, the frequency of users accessing streaming media, and the size of the storage space occupied by the project; S2: Constructing the item-user rating matrix and item-attribute matrix based on the streaming preference data; fusing the item-user rating matrix and the item-attribute matrix to obtain the item-user attribute preference matrix; S3: Calculate the comprehensive similarity of users based on the item-user rating matrix and the item-user attribute preference matrix; S4: If there are items in the item-user rating matrix that the user has not participated in rating, the K nearest neighbor users of the user are determined according to the comprehensive similarity, and the rating of the user on the item is predicted based on the ratings of the K nearest neighbor users, and the predicted rating is filled into the item-user rating matrix; S5: According to the frequency of users' access to streaming media data, users are divided into multiple levels, and the analytic hierarchy process (AHP) is used to determine the scoring weights of users at different levels; S6: Calculate the scores of all users for each project based on the user's score weight and the filled project-user score matrix, determine the final score of each project, and arrange the projects into a file list according to the final score; S7: The particle swarm algorithm is used to calculate the best cache solution based on the order of the items in the file list, the cache space of the MEC server, the size of the storage space occupied by the items, and the final score of the items.

2. According to claim 1, a method for optimizing MEC cache based on streaming media is characterized in that: The fusion of the item-user rating matrix and the item-attribute matrix comprises: p as Represents user u a For Project I k Medium Attribute B s The degree of preference, p as ∈[0,1],r ak Represents user u a For Project I k Rating, I s Represents user u a Among all the items involved in the scoring, those with attribute B s A represents a set of items, and B represents a set of attributes of the items.

3. According to the streaming media-based MEC cache optimization method of claim 1, it is characterized in that: The S3 includes: S31: Calculate the item-user attribute preference similarity using the cosine similarity formula according to the item-user attribute preference matrix; S32: Calculate the project-user rating similarity, user preference similarity and confidence according to the project-user rating matrix, and calculate the user-rating similarity according to the project-user rating similarity, confidence and user preference similarity; S33: Calculate the user comprehensive similarity based on the user-rating similarity and the item-user attribute preference similarity.

4. According to claim 3, a method for optimizing MEC cache based on streaming media is characterized in that: The S31 includes: sim ak Represents the relationship between user a and user u for item i k Medium Attribute B s The preference similarity of item i k The number of attributes in p as Represents user a's response to item i k Medium Attribute B s The degree of preference, p us Represents user u's response to item i k Medium Attribute B S degree of preference.

5. According to claim 3, a method for optimizing MEC cache based on streaming media is characterized in that: The S32 includes: S321: Calculate the item-user rating similarity based on the item-rating matrix: sim1(u,v,i) represents the similarity of the ratings of users u and v on the projects they participated in together. u represents the user, v represents another user, and P vi represents the rating of user v on item i, P ui I represents the score of item i by user u, and the set of items that user v participates in scoring is represented by I v Indicates that the set of items that user u participates in rating is represented by I u Indicates that I u ∩I v represents the common items that users u and v participate in rating, where i∈I u ∩I v ; S322: Calculate the user preference similarity based on the item-rating matrix: sim2(u,v,i) represents the similarity of rating preferences of users u and v for the projects they jointly participated in, sim2(u,v,i)∈[0,1], P ui represents the rating of item i by user u, P vi represents the rating of user v on item i, Represents the average ratings of users u and v on all participating projects, where i∈I u ∩I v ; S323: Calculate the confidence between different users based on the item-rating matrix: sim3(u,v) represents the confidence between user u and user v, I u ∩I v I represents the common items that users u and v participate in rating. u ∪I v Represents all items rated by user u or user v; S324: Calculate the user-rating similarity based on the item-user rating similarity, the user preference similarity and the confidence: sim u (u,v) represents the rating similarity between user u and user v, where C uv =I u ∩I v , represents the set of items that user u and user v jointly participate in rating, sim represents the fusion similarity of item-user rating similarity, user preference similarity and confidence, where i∈I u ∩I v .

6. The method for optimizing MEC cache based on streaming media according to claim 3, characterized in that: The S33 includes: calculating the user comprehensive similarity according to the user-rating similarity and the item-user attribute preference similarity: yes'=(1-α)yes u +αyes ak sim' represents the comprehensive similarity of users sim u Represents item-user rating similarity, sim ak represents the item-user attribute preference similarity, and α represents the proportion of item-user attribute preference similarity.

7. The method for optimizing MEC cache based on streaming media according to claim 1, characterized in that: The predicting of the user's rating of the item includes: It represents the rating prediction for the target user u who did not participate in the rating project i0. represents the average rating of user u for all projects he has participated in, U' represents the set of K nearest neighbor users most similar to the target user u, represents the rating of the neighboring user u' on the item i0, represents the average rating of all participating projects by neighboring user u', sim'(u,u') represents the comprehensive similarity between target user u and neighboring user u', and i0 represents the projects that target user u has not participated in rating.

8. The method for optimizing MEC cache based on streaming media according to claim 1, characterized in that: The S7 includes: S71: constructing an objective function of the particle swarm algorithm and a limiting function of the objective function according to the cache space of the MEC server, the size of the storage space occupied by the project, and the final score of the project; S72: according to the number of items in the file list, the corresponding number of particle groups, the corresponding spatial dimension and the corresponding number of iterations are obtained, the spatial dimension is equal to the number of items in the file list, each dimensional spatial velocity vector of each particle is randomly initialized to a real number, and each dimensional spatial position vector of each particle is randomly initialized to 0 or 1; S73: Calculate the particle position increment: in, X N represents the set of position increments of particle N in d-dimensional space, is the position increment of particle N in the d-dimensional space after the k-th iteration, represents the velocity of particle N in d-dimensional space, is a uniform random number between (-1, +1), For is a variable, and the hyperbolic tangent function takes values ​​in (-1, +1); S74: Calculate the fitness value of each particle according to the objective function and the position increment of the particle in space, and take the highest fitness value of the particle as the fitness value of the group particles; S75: Determine whether the fitness value of the particle satisfies the limiting function. When the fitness value of the particle does not satisfy the limiting function, set the fitness value of the particle to 1. When the fitness value of the particle satisfies the limiting function and the fitness value of the particle is greater than its historical best fitness value, replace the historical best fitness value of the particle with the current fitness value of the particle. When the historical best fitness value of a particle is greater than the historical best fitness value of the group particles, replace the historical best fitness value of the group particles with the historical best fitness value of the particle and record the position increment corresponding to the historical best fitness value of the current group particles as a feasible solution. S76: Update the particle speed and particle position; S77: Repeat S73-S76 until the iteration ends, output the position increment corresponding to the historical best fitness value of the group particles, determine the corresponding cache list, and obtain the best cache solution.

9. The method for optimizing MEC cache based on streaming media according to claim 8, characterized in that: The objective function of the particle swarm algorithm and the limiting function of the objective function include: max f(X1,X2,...,X N ) is the total score of all items in the cache solution, which means that in the particle swarm algorithm, max f(X1,X2,...,X N ) represents the fitness value of the particle, X N represents the set of position increments of particle N in d-dimensional space, n represents the total number of items in the file list, and w i Indicates the size of the storage space occupied by the project, V indicates the cache space size of the MEC server, and C i Represents the final score of the i-th item. When caching the i-th item in the file list, x i Equal to 1, when the i-th item in the file list is not cached x i Equal to 0, i represents the order of the items in the file list, As the objective function max f(X1,X2,...,X N ) is a limiting function.

Images