Edge caching method for perceiving user preference based on knowledge graph

Abstract

The invention requests to protect an edge caching method for perceiving user preferences based on a knowledge graph, and the method specifically comprises the steps: firstly, providing a time-sharing knowledge graph attention network to perceive user preferences, and splicing a user content interaction graph and a content entity knowledge graph into a collaborative knowledge graph; a time sequence relation and a sharing relation are introduced to reconstruct an interaction relation between a user and content, entity high-order information spreading is performed by adopting double attention spreading layers, results of different layers are aggregated by utilizing a gating mechanism, and then a user preference score is calculated through an inner product; calculating the popularity of the content in the user group in combination with the user activeness and the user preference; and finally, considering edge cache capacity limitation, constructing a cache optimization problem with maximized content popularity, and determining edge cache content. According to the method, user preference learning is enhanced semantically by using the knowledge graph, the preference perception accuracy is improved from the time dimension through the sequential relationship, and then the cache content is determined based on the user preference, so that the edge cache better meets the user demand, and the hit rate is improved.

Description

Technical Field

[0001] The present invention relates to the field of edge caching, and in particular to an edge caching method based on user preference. Background Art

[0002] In the current digital age, cloud computing has become the mainstream way to store and process data. However, with the popularity of IoT devices and mobile applications, the amount of data is growing exponentially, and traditional cloud computing architecture faces bandwidth bottlenecks and high latency problems. In order to meet these challenges, edge computing came into being. Edge computing can effectively reduce latency, improve data processing efficiency and user experience by deploying computing and storage resources on the edge nodes of the network.

[0003] In the framework of edge computing, edge caching significantly reduces the distance and time of data transmission by storing data close to the user, thereby improving the system's response speed and bandwidth utilization. In order to formulate edge caching strategies more effectively, understanding user preferences is key, because user preferences will directly or indirectly affect the selection of cached content. User preferences are not only reflected in access frequency and trends, but also involve deeper semantic understanding. By analyzing user preferences, it is possible to predict which content is most popular with users, analyze the overall preference of user groups in the edge network for content, and obtain the overall popularity of the content, so as to give priority to caching these popular content.

[0004] By revealing the semantic relationship between content, the knowledge graph can provide a deeper understanding of user interest preferences. It can not only identify the implicit associations between users and different content, but also help discover users' potential areas of interest. By combining users' access behavior with the semantic structure of content, the knowledge graph can provide a more comprehensive user behavior analysis. This semantic level of understanding allows edge caching strategies to more accurately predict user needs, thereby optimizing caching decisions. Therefore, how to use knowledge graphs to semantically perceive user interest preferences and formulate edge caching strategies based on user preferences is a key issue that needs to be solved in edge computing.

[0005] Therefore, the present invention proposes an edge caching method based on knowledge graph to perceive user preferences.

[0006] After searching, the application publication number CN118748643A belongs to the edge cache method based on user preference. The method includes providing an edge cache placement optimization method and system based on user preference perception. The specific steps are as follows: based on the mobile edge network, determine the overall access delay of the user's requested content; obtain the user's historical request record, use a deep learning network to predict the number of user-requested content in the next time slot, and combine the activity level of heterogeneous users to obtain the probability distribution of user preferences for content; based on the probability distribution of user preferences for content, cluster users according to preference similarity, divide highly related users into the same cluster, and design a cache strategy for each cluster; transform the cache placement problem into a delay reduction maximization problem, maximize the overall delay reduction under the constraints of cache capacity and delay threshold, use a genetic algorithm to optimize the delay reduction maximization problem, and iterate the best individual as the global optimal cache placement strategy; according to the global optimal cache placement strategy, execute the placement of cache content, and dynamically adjust the cache content according to user requests to maximize the cache hit rate.

[0007] This patent aims to use deep learning networks to perceive user preferences for content directly from user historical request records, that is, to understand user interest preferences only from user access patterns, lacking understanding from a semantic level. Unlike this patent, this patent uses a collaborative knowledge graph that combines a user content interaction graph with a content entity knowledge graph to learn user interest preferences from a deeper semantic level, while introducing temporal relationships and sharing relationships to further learn the representation of user interests from a time dimension. In addition, this patent also uses a fusion of multiple features to represent user activity in order to better perceive changes in user activity. Summary of the invention

[0008] The present invention aims to solve the above problems of the prior art. An edge caching method based on knowledge graph perception of user preferences is proposed. The technical solution of the present invention is as follows:

[0009] An edge caching method based on knowledge graph perception of user preferences comprises the following steps:

[0010] S1. Combining the user content interaction graph and the content entity knowledge graph, the time-sharing knowledge graph attention network is used to predict the user's content interest preference;

[0011] S2. Calculate user activity by comprehensively calculating the user activity based on the frequency domain features and time domain features of the user's historical requests;

[0012] S3, calculate the content popularity, based on the user's content interest preference and user activity to calculate the current overall popularity of the content;

[0013] S4. Determine the optimization goal of maximizing content popularity, construct the content cache optimization problem considering the edge cache capacity limitation, and obtain the final cache content solution.

[0014] Furthermore, in step S1, combining the user content interaction graph and the content entity knowledge graph, using the time-sharing knowledge graph attention network to predict the user's content interest preference specifically includes:

[0015] S11. Introducing temporal relationship r into user content interaction graph t and shared relations s , combined with the content entity knowledge graph, the TransR method is used in the embedding layer to learn the vector representation of entities and relations in the graph;

[0016] S12. The user content interaction graph with temporal relations and the user content interaction graph with sharing relations are respectively spliced ​​with the content entity graph into knowledge graphs with different relations. The attention mechanism and graph neural network are combined to use two parallel attention propagation layers to learn high-level representations of entities under different relations respectively, and then the gated aggregation mechanism is used to aggregate entity representations of different orders.

[0017] S13. Aggregate entity representations of different orders using an aggregation function to obtain the final information embedding of each entity. Perform an inner product on the entity information embedding representations of the user and content to obtain the user's interest preference score for the content.

[0018] Furthermore, in step S2, the user activity is calculated based on the frequency domain features and time domain features of the user's historical requests, specifically including:

[0019] S21. Evaluate the user's request frequency characteristics based on the user's historical request times;

[0020] S22. Evaluate the user's request time domain characteristics based on the number of time slots spanned by the user's historical requests;

[0021] S23. Comprehensively calculate the user's activity based on the frequency domain features and the time domain features.

[0022] Furthermore, in step S3, the content popularity is calculated, and the current overall popularity of the content is calculated based on the user's content interest preference and user activity, specifically including:

[0023] S31. Calculate user similarity based on the user's content interest preference, and obtain a set of similar users of the user based on a similarity threshold;

[0024] S32. Calculate the overall current popularity of the content based on the user's content interest preference, user activity, and similar user collection;

[0025] S33. Calculate the latest popularity of the content based on the historical popularity and current popularity of the content.

[0026] Furthermore, in step S4, the optimization goal of maximizing content popularity is determined, and the content cache optimization problem is constructed considering the edge cache capacity limitation to obtain the final cache content solution, which specifically includes:

[0027] S41. Determine maximizing content popularity as the optimization goal of cache content, and construct a content cache optimization problem considering the edge cache capacity limitation;

[0028] S42. Use a dynamic programming algorithm to solve the optimization problem and obtain the final cache content set.

[0029] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, an edge caching method based on knowledge graph-aware user preference as described in any one of the items is implemented.

[0030] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an edge caching method based on knowledge graph-aware user preference as described in any one of the items.

[0031] A computer program product comprises a computer program, wherein when the computer program is executed by a processor, the edge caching method based on knowledge graph-aware user preference as described in any one of the items is implemented.

[0032] The advantages and beneficial effects of the present invention are as follows:

[0033] 1. In step S1 of the present invention, the rich semantic information of the knowledge graph is used to supplement the edge content information, the user content interaction graph and the content entity knowledge graph are spliced ​​into a collaborative knowledge graph, and the proposed time-sharing knowledge graph attention network is used to more accurately capture the user's interest preferences in edge scenes from a semantic level, thereby improving the perception accuracy of user preferences.

[0034] 2. In step S1 of the present invention, by introducing the timing relationship r t and shared relations s Reshape the interactive behavioral relationship between users and content, use a double-layer attention propagation layer and a gating mechanism to aggregate information propagation under different interactive behavioral relationships, and more effectively capture changes in user interests from a temporal dimension. This overcomes the problem of lack of temporal dynamics and neglect of the similarity of user interests changing over time in traditional collaborative filtering methods that only rely on user-content interactive relationships.

[0035] 3. In step S2, the present invention combines the frequency characteristics and time series characteristics of the user's interactive behavior history to comprehensively measure the activity of marginal users, making the user activity evaluation more accurate and practical.

[0036] 4. In step S3, the present invention introduces a cosine similarity calculation method with norm difference penalty to more accurately measure the differences in preference strength between users, thereby effectively solving the problem of excessive similarity caused by different interest preference strengths. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flow chart of the overall preferred embodiment provided by the present invention;

[0038] Figure 2 It is a time-sharing knowledge graph attention network model diagram of the present invention;

[0039] Figure 3 This is an example diagram of the collaborative knowledge graph of the present invention. DETAILED DESCRIPTION

[0040] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0041] The technical solution of the present invention to solve the above technical problems is:

[0042] The present invention provides an edge caching method based on knowledge graph perception of user preferences, characterized by comprising the following steps:

[0043] S1. Combining the user content interaction graph and the content entity knowledge graph, the time-sharing knowledge graph attention network is used to predict the user's content interest preference;

[0044] S2. Calculate user activity, based on the frequency domain features and time domain features of the user's historical requests;

[0045] S3. Calculate the content popularity, and calculate the current overall popularity of the content based on the user's content interest preference and user activity;

[0046] S4. Determine the optimization goal of maximizing content popularity, construct the content cache optimization problem considering the edge cache capacity limitation, and obtain the final cache content solution.

[0047] In this embodiment, in step S1, the user content interaction graph and the content entity knowledge graph are combined to use the time-sharing knowledge graph attention network to predict the user's content interest preference, which specifically includes the following steps:

[0048] S11. Embedding layer learns entity vector representation

[0049] The input of the time-sharing knowledge graph attention network is the recent user content interaction graph G1 and the content entity knowledge graph G2, and the output is the user preference score.

[0050] In recommendation systems, collaborative filtering methods are often used to predict user preferences for content, thereby achieving content recommendation. Collaborative filtering assumes that if two users have interactive behaviors on the same content, then the two users have certain similarities in interest preferences. User preferences are not only affected by the user's own interactive behaviors, but are also closely related to the behavioral patterns of other users with similar interests. However, the similarity of interest preferences between users may also be affected by the time dimension. For example, if two users interact with the same content in different time periods, although the two users may have certain similarities in terms of behavioral patterns, their similarity may be lower in terms of time dimension. On the contrary, when two users interact with the same content in the same time period, they can be considered to have higher similarity in the time dimension, which indicates that they may show stronger interest preferences for the same or the same type of content in the same time period.

[0051] The user content interaction graph is a bipartite graph consisting of all the interactions of all the users on the edge network over a certain period of time in the past. First, in order to better capture the user’s interest preferences from the time dimension, the temporal relationship r is introduced into the user content interaction graph. t , recording interactions in different time periods as different relationships, thus representing the user content interaction graph as a knowledge graph containing different temporal relationships In addition, in order to maintain the analysis of user interest preferences from the behavioral pattern in the collaborative filtering method, the sharing relationship r is introduced in the user content interaction graph. s , understand the user content interaction graph as a knowledge graph with unified interaction relationships To maintain the information transmission of user and content interaction behavior patterns.

[0052] The content entity knowledge graph is constructed by combining an open and shared knowledge graph database. For each content, some matching rules are used, such as extracting keywords such as content title and text, or other information, to map the content to a knowledge graph database such as Freebase, so that each content is associated with the knowledge entity of the knowledge base to form a triple and build the content knowledge graph G2.

[0053] Then, the user content interaction graphs with different relationships are respectively spliced ​​with the content entity knowledge graph to form a collaborative knowledge graph G s and G t , where G s The relationship between users and content in G is a sharing relationship. t The relationship between user and content interaction is a temporal relationship.

[0054] The time-sharing knowledge graph attention network is divided into three layers: embedding layer, hierarchical attention propagation layer, and aggregation layer. In the embedding layer, the TransR method is used to represent the triple information in the knowledge graph to learn the vector embedding of entities and relations. The TransR method represents a triple (h, r, t) as in and It is the head entity h and tail entity e t The projection on relation r has a credibility score of:

[0055]

[0056] in, is the transformation matrix of r, which projects the entity from the d-dimensional entity space to the k-dimensional relationship space.

[0057] With the following loss function L KG To distinguish the relative order of valid and invalid triplets:

[0058] L KG =L KG (r t )+L KG (r s ) (2)

[0059]

[0060] in, (h,r s ,t′) is an invalid triple formed by randomly replacing one entity in a valid triple. Similarly, (h,r t ,t′) is also an invalid triple formed by randomly replacing one entity in the valid triple. Because there are two different relationships between users and content, the loss functions under different relationships are calculated separately.

[0061] S12. Attention propagation layer enhances entity information aggregation

[0062] In the hierarchical attention propagation layer, the attention mechanism is combined with the graph neural network to capture user similarity features in the user content interaction graph and deep semantic associations in the content entity knowledge graph, thereby perceiving user interest preferences.

[0063] In collaborative filtering methods, connectivity based on behavioral patterns is usually used to predict user preferences. For example, in the user content interaction graph, there is a set of data: It is believed that u1 has a certain interest preference for i2. Because users u1 and u2 are connected to i1 through relationship r1, it is believed that u1 and u2 are similar users, and u2 is connected to i2 through relationship r1, so u1, as a similar user of u2, may also show interest preference for i2. The relationship here can be a simple sharing relationship that only represents interactive behavior in collaborative filtering, or a time series relationship divided by time period. For example, in In , assuming that r1 and r2 are different temporal relationships divided by time periods, here u1 and u2 both interact with i1 in the same time period, indicating that u1 and u2 have more similar interests and preferences, so u1 may show interest in i2 in the time period corresponding to the relationship r2. However, if In , u1 and u2 interact with i1 in different time periods, indicating that the preference similarity between u1 and u2 is low, so u1 may show a lower interest preference for i2 in the time period corresponding to the relationship r2.

[0064] In addition, the added content entity knowledge graph provides additional attributes or semantic expansion for the content, thereby further analyzing user preferences through attribute-based high-order connectivity. In this graph connection, it can be assumed that u1 has a certain interest preference for i2, because u1 has interacted with i1, and i1 and i2 are connected to e1 through the same relationship r3, that is, i1 and i2 have the same attribute e1.

[0065] Therefore, by analyzing the high-order connectivity of the graph, we can better learn the user's preference score for content. This method recursively learns the information propagation embedding of entities at different orders through high-order connectivity, and uses the idea of ​​graph attention network to generate cascade propagation attention weights to reveal the importance of this connectivity. In order to simulate the information propagation process between nodes in the collaborative knowledge graph, given an entity h, let N h = {(h, r, t)|(h, r, t)∈G} is the set of all triples of which it is the head entity. To characterize the first-order connectivity structure of entity h, we calculate N h The linear combination of derives the first-order representation of entity h:

[0066] e Nh =∑ (h,r,t)∈N π(h,r,t)e t (5)

[0067] Where π(h,r,t) is the attenuation factor that controls each propagation of a triple edge (h,r,t), indicating the amount of information propagated from the tail entity t to the head entity h through the relation r.

[0068] Represent π(h,r,t) through the relational attention mechanism:

[0069] π(h,r,t)=(W r e t ) T tanh((W r e h +e r )) (6)

[0070] Where tanh is a nonlinear activation function, which makes the attention score depend on the relationship e h To e r That is to say, the closer the head entity is to the tail entity, the more information is transmitted.

[0071] Then the softmax function is used to normalize the coefficients of all triplets connected to h:

[0072]

[0073] The original e h After information dissemination Aggregated by aggregation function into e containing first-order connectivity information propagation h express:

[0074]

[0075] Where LeakyReLU is the activation function, are trainable weights and ⊙ are element-wise products.

[0076] Then multi-hop relation propagation can explore the high-order connection information of entities by stacking more propagation layers. For step l, the entity is recursively represented as:

[0077]

[0078] Two parallel attention propagation layers are used here to process the temporal relationship r t The temporal relationship layer of the user content interaction graph and content entity knowledge graph and the processing of shared relations r s The shared relationship layer of the user content interaction graph and the content entity knowledge graph. The layer containing the temporal relationship can learn the similarity between users from the time dimension, so as to better capture the interests of users at different time points and obtain a better representation of user preferences. The layer containing the shared relationship captures the interest representation of user behavior patterns according to the idea of ​​collaborative filtering. The two layers of attention propagation layers will respectively obtain the high-order representation of entity h and The gated aggregation mechanism is used to aggregate the representations of entity information propagated at different orders in the two layers. First, the gating vector g used to control the weights of different layers in the final gated aggregation result is calculated. e :

[0079]

[0080] in, Indicates that and For splicing, W g and b g They are all trainable parameters, and sigmoid() is the activation function.

[0081] Then, the gating vector is used to combine the features of the results of different layers to obtain a high-level representation of the final entity:

[0082]

[0083] S13. Aggregation layer calculates user preference scores

[0084] After executing L layers, each user u has multiple representations {e u (1) ,...,e u (L)}, each content also corresponds to multiple representations {e c 1 ,...,e c (L)}, the outputs of different layers emphasize connectivity information of different orders, so a layer aggregation mechanism is used to connect the representation of each step into a vector:

[0085] e u * =e u (0) ||...||e u (L) (13)

[0086] e c * =e c (0) ||...||e c (L) (14)

[0087] Here, || is a connection operation, which can not only perform embedding propagation operations to enrich the initial embedding, but also control the propagation strength by adjusting L, and finally obtain representations after information propagation of different orders.

[0088] Finally, the inner product of the user and content representation is performed to obtain the user's interest score in the content:

[0089]

[0090] For user u, its interest score for all content is represented as vector I u .

[0091] Furthermore, in step S2, the user activity is calculated based on the frequency domain features and time domain features of the user's historical requests, specifically including:

[0092] S21. Evaluate the user's request frequency domain characteristics based on the number of historical requests. Here, the total number of requests is used to measure the frequency domain characteristics:

[0093] AF u =∑ c∈Contents q u,c (16)

[0094] Contents represents the content collection, q u,c Represents the number of interactions between user u and content c.

[0095] S22. Evaluate the user's request time domain characteristics based on the number of time slots spanned by the user's historical requests. The number of time slots in which the user interacts with the content reflects the breadth of the user at the time level. The more different time slots spanned by the interaction, the more active the user is at the time level. The time domain characteristics are expressed as follows:

[0096] AT u =∑ t∈Slots r u,t (17)

[0097] Slots represents a set of time slots, which are generally divided into one time slot per hour. They can also be flexibly divided according to actual application scenarios. If user u has an interaction history in time slot t, then r u,t =1, otherwise r u,t =0.

[0098] S23. Use the multiplication rule to fuse the frequency domain features and time domain features to obtain the user's activity:

[0099] A u =AF u ×AT u (18)

[0100] Furthermore, in step S3, the current overall popularity of the content is calculated based on the user's content interest preference and user activity, specifically including:

[0101] S31. Calculate similar user sets

[0102] In order to improve the cache hit rate and make the cached content more consistent with the overall preference trend of edge users, the content popularity calculated by user preferences is used to characterize the overall preferences of all users. Considering that the contribution of a single user's preference to content popularity will be affected by its similar users, cosine similarity is used to calculate the similar user set of each user based on the user's content interest preference. However, cosine similarity tends to measure the similarity of user interests in a certain direction and ignores the difference in interest intensity. Therefore, a norm difference penalty term is added to improve cosine similarity:

[0103]

[0104] Where λ is the penalty factor, which controls the strength of the penalty term. Indicates I u and I v The difference in norm.

[0105] Since the dimension is too high, the results calculated using cosine similarity will be similar and the value is close to 1. Therefore, the maximum-minimum normalization is used to reshape the similarity values ​​between users:

[0106]

[0107] Among them, Users is a collection of users.

[0108] Next, the similar user set of user u is obtained according to the similarity threshold κ:

[0109] S u ={v|sim norm (u,v)>κ,v∈Users} (21)

[0110] S32. Calculate content popularity

[0111] Calculate the popularity contribution score of user u to content c based on the user's content interest preference and user activity:

[0112]

[0113] Then, using the user's contribution to the popularity of the content and the indirect influence of similar user sets, the following formula is used to calculate the current overall popularity of a certain content in the user group:

[0114]

[0115] Among them, β is a value that controls the degree of indirect influence of similar users on user interests. The current popularity value is calculated based on recent user content interaction behavior, reflecting the popularity of the content in the near future, while the historical popularity reveals the popularity trend of the content among user groups over a long period of time in the past. In order to more accurately characterize the popularity trend of content among user groups and balance short-term fluctuations with long-term trends, the final popularity representation of the content can be updated by combining the current popularity and historical popularity:

[0116] p(c)=α·p old (c)+(1-α)·p new (c) (24)

[0117] Among them, p(c) represents the final popularity of content c, p old (c) represents the historical popularity of the content, and α is a factor that controls the degree of expression of new and old popularity.

[0118] Furthermore, in step S4, the optimization goal of maximizing content popularity is determined, and the content cache optimization problem is constructed considering the edge cache capacity limitation to obtain the final cache content solution, which specifically includes:

[0119] S41. Determine maximizing content popularity as the optimization goal of cache content, and consider the edge cache capacity limitation to construct the content cache optimization problem:

[0120]

[0121] Where f(c) indicates whether content c is cached, if yes, it is 1, otherwise it is 0. size(c) indicates the size of content c.

[0122] S42. Use a dynamic programming algorithm to solve the optimization problem and obtain the final cache content set.

[0123] The present invention is further described below with specific examples:

[0124] Assume that a total of 4 pieces of content are stored in the cloud under a certain cloud-edge architecture, and there are 5 users in a certain edge network. The content interaction history of users in the past day is as follows:

[0125] User ID Interactive content ID Interaction time 1 1 2024.11.20 01:30:00 1 2 2024.11.20 21:47:45 2 2 2024.11.20 08:11:12 2 3 2024.11.20 17:13:56 3 3 2024.11.20 15:09:12 4 4 2024.11.20 23:22:31 5 1 2024.11.20 16:28:36 5 3 2024.11.20 16:45:36

[0126] The content entity knowledge graph is constructed based on content attributes, keywords, features, etc., combined with the actual knowledge base. Here, it is assumed that Amazon-Book is used as the content data set, and then each content is mapped to the entity in the Freebase knowledge database according to the title match to establish a content entity knowledge graph. In addition, the time of user content interaction is divided into multiple segments to represent the temporal relationship between user and content interaction, and then a collaborative knowledge graph containing the temporal relationship is constructed. For the sake of simplicity, a day is divided into 4 time periods: 00:00:00-06:00:00; 06:00:00-12:00:00; 12:00:00-18:00:00; 18:00:00-24:00:00, such as Figure 3 .

[0127] The user content interaction graph and content entity knowledge graph are input into the time-sharing knowledge graph attention network to obtain the user content interest preference. All calculated user preferences are expressed in a matrix as follows:

[0128]

[0129] Then, the time slots are divided according to the rules of dividing the temporal relationship of the user content interaction graph, and the activity of each user is obtained according to the user activity calculation formula. For example, for user u1, there are 2 interaction records in the past, and its frequency domain feature is 2. According to the time slot division, these two interaction behaviors are located in different time slots, so their time domain feature is 2, and then the activity of user u1 is calculated as 4 by combining the frequency domain feature and the time domain feature. The calculation of other users is similar, and finally the activity of the five users is: 4, 4, 1, 1, 2.

[0130] According to the user preference matrix calculated above, the similarity between users is calculated using formulas (19) and (20):

[0131]

[0132] The user similarity threshold is set to 0.7, and the similar user set of each user is calculated by formula (21). Then, the overall popularity of the content on the edge is calculated based on user preference, user activity, and similar user set. The results are 19.160, 18.406, 19.633, and 16.429 respectively. The popularity of the content in the user group in the past time is not considered here.

[0133] Assume that the file sizes of the four contents are [10, 9, 8, 7], and the edge cache capacity is limited to 25. Then the caching problem is converted into an optimization problem of maximizing the overall popularity of the cached content, and the final cached content set is [1, 3, 4] solved by the dynamic programming algorithm.

[0134] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0135] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0136] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0137] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. An edge caching method based on knowledge graph to perceive user preferences, characterized in that: The following steps are involved: S1. Combining the user content interaction graph and the content entity knowledge graph, the time-sharing knowledge graph attention network is used to predict the user's content interest preference; S2. Calculate user activity by comprehensively calculating the user activity based on the frequency domain features and time domain features of the user's historical requests; S3, calculate the content popularity, based on the user's content interest preference and user activity to calculate the current overall popularity of the content; S4. Determine the optimization goal of maximizing content popularity, construct the content cache optimization problem considering the edge cache capacity limitation, and obtain the final cache content solution.

2. According to claim 1, the edge caching method based on knowledge graph perception of user preferences is characterized in that: In step S1, combining the user content interaction graph and the content entity knowledge graph, using the time-sharing knowledge graph attention network to predict the user's content interest preference specifically includes: S11. Introducing temporal relationship r into user content interaction graph t and shared relations s , combined with the content entity knowledge graph, the TransR method is used in the embedding layer to learn the vector representation of entities and relations in the graph; S12. The user content interaction graph with temporal relations and the user content interaction graph with sharing relations are respectively spliced ​​with the content entity graph into knowledge graphs with different relations. The attention mechanism and graph neural network are combined to use two parallel attention propagation layers to learn high-level representations of entities under different relations respectively, and then the gated aggregation mechanism is used to aggregate entity representations of different orders. S13. Aggregate entity representations of different orders using an aggregation function to obtain the final information embedding of each entity. Perform an inner product on the entity information embedding representations of the user and content to obtain the user's interest preference score for the content.

3. According to claim 1, the edge caching method based on knowledge graph perception of user preferences is characterized in that: In step S2, the user activity is calculated based on the frequency domain features and time domain features of the user's historical requests, specifically including: S21. Evaluate the user's request frequency characteristics based on the user's historical request times; S22. Evaluate the user's request time domain characteristics based on the number of time slots spanned by the user's historical requests; S23. Comprehensively calculate the user's activity based on the frequency domain features and the time domain features.

4. According to claim 1, the edge caching method based on knowledge graph perception of user preferences is characterized in that: In step S3, the content popularity is calculated, and the current overall popularity of the content is calculated based on the user's content interest preference and user activity, specifically including: S31. Calculate user similarity based on the user's content interest preference, and obtain a set of similar users of the user based on a similarity threshold; S32. Calculate the overall current popularity of the content based on the user's content interest preference, user activity, and similar user collection; S33. Calculate the latest popularity of the content based on the historical popularity and current popularity of the content.

5. According to claim 1, the edge caching method based on knowledge graph perception of user preferences is characterized in that: In step S4, the optimization goal of maximizing content popularity is determined, and the content cache optimization problem is constructed considering the edge cache capacity limitation to obtain the final cache content solution, which specifically includes: S41. Determine maximizing content popularity as the optimization goal of cache content, and construct a content cache optimization problem considering the edge cache capacity limitation; S42. Use a dynamic programming algorithm to solve the optimization problem and obtain the final cache content set.

6. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the edge caching method based on knowledge graph-aware user preference as described in any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the edge caching method based on knowledge graph-aware user preference as described in any one of claims 1 to 5 is implemented.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the edge caching method based on knowledge graph-aware user preference as described in any one of claims 1 to 5 is implemented.

Images