An edge data caching method

Abstract

This invention relates to an edge data caching method, belonging to the field of the Internet of Things (IoT). The method includes: data collection; user search record storage, including edge server storage of user historical query content and search feedback data; edge inference, including a hotspot entity identification method, sequentially constructing an individual interest prediction model and a group interest prediction model, and identifying hotspot entities based on the degree of interest of group users; edge caching, including prioritizing the caching of hotspot entity information through a designed edge caching strategy; content replacement: completing the cache replacement of entity information through a designed content replacement strategy; user search: the edge server responds to user search requests directly or through communication with sensing devices. This invention can effectively improve system caching performance and enhance the user search experience.

Description

Technical Field

[0001] This invention belongs to the field of the Internet of Things and relates to an edge data caching method. Background Technology

[0002] In recent years, the large-scale deployment of sensing devices has led to the rapid development of the Internet of Things (IoT), with massive amounts of information flooding cyberspace to provide various application services for IoT users. Due to the extremely limited storage resources of local sensing devices, the common practice is to upload entity information to a remote central server with powerful computing and storage resources. However, IoT entity information is dynamically changing, and the long transmission latency between the central server and the user cannot guarantee the freshness of the information. This results in a significant loss of freshness in the entity information obtained by the user, impacting the user's search experience. To address these issues, edge caching technology has emerged.

[0003] Edge caching, as an emerging solution in edge computing, can utilize its storage resources to respond to user needs locally, thereby avoiding frequent end-to-end communication and reducing long-distance communication between users and remote central servers. This frees up limited terminal resources while effectively ensuring the timeliness of entity information. However, existing edge caching strategies still face several challenges due to practical constraints. First, while edge servers possess considerable storage resources, their cache capacity is ultimately limited, necessitating entity filtering and caching mechanism design within this capacity constraint to improve system cache efficiency. Second, with the increasing demand for entity information from IoT users, a user-centric edge caching approach is urgently needed to provide a better service experience. Against this backdrop, this paper aims to design an efficient entity information caching strategy from the perspective of user preferences to meet users' search requests for entity information.

[0004] Currently, there are some research works on edge caching. Li L, Xu Y, and Yin J, in "Deepreinforcement learning approaches for content caching in cache-enabled D2D networks" [in IEEE Internet of Things Journal, vol. 7, no. 1, pp. 544-557, 2020], studied the joint optimization of cached content placement and delivery strategies in cache-enabled D2D networks. They used echo-state networks and long short-term memory networks to predict user mobility and content popularity, thereby determining cached content and cache location. Song J, Sheng M, and Tony Q, in "Learning-based content caching and sharing for wireless networks" [in IEEE Transactions on Communications, vol. 7, no. 1, pp. 4309-4324, 2017], sought an optimal caching strategy by jointly optimizing the costs of content caching, content sharing, and content retrieval in base stations using a multi-armed caching machine approach when the popularity distribution is unknown. Cui L., in "CREAT: Blockchain-assisted Compression Algorithm of Federated Learning for Content Caching in Edge Computing" [in IEEE Internet of Things Journal], focuses on user preferences while considering user data privacy. He introduces a federated learning framework to train a user preference model on the local terminal, utilizes an autoencoder to mine latent features of users and files, and combines this with user social relationships through hybrid filtering to generate a file recommendation list for the user. While this method considers user preferences to improve caching performance for user search requests, it neglects the evolution of user interests over time—specifically, long-term and short-term interests—and fails to consider the varying degrees of influence of each historical search record on user interest perception. This results in poor accuracy in perceiving user preferences, impacting entity popularity identification centered on user experience and ultimately affecting caching performance.

[0005] The multidimensional attributes of IoT entities and the search context of users lead to differentiated preferences for entities, resulting in each physical entity having its own popularity. For example, a "coffee shop" in a bustling area will be searched more frequently than a "forest" in a sparsely populated area. Prioritizing the caching of entities with high popularity can meet the needs of more users, improve the user service experience, and maximize the system caching performance. Summary of the Invention

[0006] In view of this, the first objective of this invention is to provide a method for identifying hot entities. This method calculates entity popularity and identifies hot entities from the perspective of user search experience. First, it mines user interests and preferences based on historical search records of IoT users (including search requests, clicked entities, feedback evaluations, entity ratings, etc.), constructs a personal interest prediction model, and uses this model to predict the list of entities of interest to individual users. Then, it utilizes the concept of "social computing" to construct a group interest prediction model to obtain the popularity of entities. Finally, it sets a threshold to distinguish between hot and non-hot entities, achieving hot entity identification and solving the problem that the limited accuracy of existing entity popularity identification methods affects the final user experience.

[0007] The second objective of this invention is to provide an edge data caching method. Based on the obtained entity popularity, the edge caching strategy is designed by considering the constraints of entity information memory size and system cache capacity, so as to maximize the system cache hit rate and effectively meet the high-quality service requirements for user search requests.

[0008] To achieve the first objective mentioned above, the present invention provides the following technical solution:

[0009] A hotspot entity recognition method specifically includes the following steps:

[0010] S1: Store user search history: The edge server stores the search history of each user u within the managed area. in Indicates user u in t j The entity e was searched at all times. j ;

[0011] S2: Building an Individual Interest Model: The edge server builds an individual interest prediction model based on the historical request records of users searching within the same area, and uses this model to predict each user's level of interest in unknown entities. Get the entity interest list of individual user u Where N represents the total number of entities;

[0012] S3: Constructing a group interest model: Based on the individual interest prediction model, analyze group users By performing joint reasoning on the entity interest list, we can obtain the entity interest list of group users. in The entity e is represented i Popularity, where M represents the total number of users;

[0013] S4: Identify Hot Entities: Analyze Entity Popularity Set a threshold ρ to distinguish between hot and non-hot entities. If it is a hot entity, then it is a hot entity; otherwise, it is a non-hot entity.

[0014] Further, in step S2, a personal interest model is constructed. This specifically includes analyzing the search history of users within the edge server's management area, considering individual users' long-term and short-term preferences, as well as the differentiated relevance of historical requests, to accurately predict the degree of interest of individual users in entities. This specifically includes the following steps:

[0015] S21: Extracting high-dimensional hidden features h of each user's short-term search history based on a Long Short-Term Memory (LSTM) network model. I(u) ;

[0016] S22: Integrating attention mechanisms to determine the impact of high-dimensional hidden features on users' unknown entity interests Weighted fusion operations are performed on high-dimensional hidden features to form the short-term preferences of individual users.

[0017] S23: Extract high-dimensional hidden features from users' long-term search history using the Asymmetric Singular Value Decomposition (Asymmetric-SVD) method, and use these features as long-term preferences for individual users.

[0018] S24: Individual user interest preferences are a comprehensive reflection of long-term and short-term interest preferences. The final individual user interest preferences are obtained by adaptively combining the two preferences through a multilayer perceptron (MLP). in W is the weighting factor, σ(·) is the sigmoid function, and W f ,b f The model learns parameters for the concat(·) method of a fully connected network; then, it calculates the parameters of individual user u for unknown entity e. i level of interest

[0019] S25: Optimize and train the model from steps S21-S24 using stochastic gradient descent (SGD), with the training loss function being... in Let λ be the predicted user preference value and θ be the actual user preference value, respectively. Let λ||Θ||2 be the regularization term, λ be the weighting factor, and Θ be all model learning parameters. Finally, the trained model is used to predict an individual user's interest list for N entities.

[0020] Furthermore, step S21 specifically includes: the LSTM model is constructed using a forgetting gate f k =σ(x k W f +h k-1 U f +b f ), Input gate i k =σ(x k W i +h t-1 U i +b i ), output gate o k =σ(x k W o +h k-1 U o +b o ), cell state c k =f k ⊙c k-1 +i k ⊙φ(x k W c +h k-1 U c +b c ) and hidden layer h k =o k ⊙φ(c k Composed of ) where W f W i W o U f U i U o b represents the weight parameters in the LSTM network model. f ,b i ,b o ,b c Here are the bias parameters in the LSTM model; σ(·) and φ(·) are the sigmoid and tanh activation functions, respectively; each short-term search history I(u) of the user is used as the LSTM input x. k The output of the hidden layer is used as the extracted high-dimensional hidden feature h. I(u) .

[0021] Furthermore, step S22 specifically includes: user u in t j Entity e being searched at any time j For unknown entity e p Predictions have differentiated impacts; an attention mechanism is introduced to calculate the relevance weights for each interaction history. Through the attention score a k Adjusting cell state in LSTM c k and hidden layers For the adjusted high-dimensional hidden vector The user's short-term preferences are obtained by weighted averaging, and are represented as follows: Among them W x , Here, k represents the sequence number corresponding to the elements in set I(u) when iterating through them sequentially, and k∈|I(u)|.

[0022] Furthermore, step S23 specifically includes: long-term search history can reflect an individual user's long-term stable interest in entities, and the user's long-term preferences are represented as... This can be obtained from its long-term search history of entities. This indicates the search history for each user. The weighting factor, The feature mapping representation of the input data. v l ,b r These are the parameters learned by the model.

[0023] Furthermore, step S3 specifically includes the following steps:

[0024] S31: According to user u x and u y User interest list and Get user u x and u y similarity sim(u x ,u y )for:

[0025]

[0026] S32: Based on user u x With all other users u y Overall similarity sim(u x ,u) Calculate user u x Similarity to groups

[0027] S33: Due to overall similarity Higher user preference for candidate entities can more accurately reflect entity popularity and obtain individual user interest weights.

[0028] S34: Computing entity e i Popularity M represents the total number of users.

[0029] To achieve the second objective mentioned above, the present invention provides the following technical solution:

[0030] An edge data caching method specifically includes the following steps:

[0031] 1) Data collection: Entity data upload: Sensing devices periodically observe entity status data and selectively upload it; User search record storage: Edge servers store users' historical query content and their search rating data;

[0032] 2) Edge reasoning: Calculate the popularity of an entity based on the hotspot entity identification method and determine whether the entity is a hotspot entity. If it is, the hotspot entity has the right to be cached first.

[0033] 3) Edge caching: When the total size of all hot entity information is less than the system cache capacity, all hot entity information is cached on the edge server. Non-hot entities are actively cached according to their popularity and size until the edge server cache capacity is full. When the total size of all hot entity information is greater than the system cache capacity, hot entities with higher popularity are cached first until the edge server cache capacity is full.

[0034] 4) Content replacement: The edge server periodically infers the interests of individual users and groups of users based on their search history, and updates the results of entity popularity and hot entity identification; at the same time, it replaces the cached entity information according to the edge caching strategy in step 3).

[0035] 5) User Search: When a user sends a search request, the edge server quickly determines whether the user is searching for a hot entity and its cache location. If it is cached on the edge server, the result is returned directly to the user; otherwise, the edge server sends the user's search request to the terminal sensing device, which collects the data and returns it to the user through the edge server.

[0036] The beneficial effects of this invention are as follows:

[0037] (1) The hot spot entity identification method proposed in this invention can obtain accurate entity popularity by mining the long-term and short-term interest preferences of individual users and constructing the interest of group users for joint reasoning. This method has strong universality and generalization ability in mining deep features and realizing prediction mechanism. Furthermore, the hot spot entity identification method can be deployed on edge servers, cloud servers, local gateways, etc., and has a wide range of applicability.

[0038] (2) The edge data caching method proposed in this invention designs a priority caching mechanism based on the popularity of entities, which not only ensures the caching performance of the system but also improves the user service experience.

[0039] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0041] Figure 1 A model diagram of an edge data caching system provided by this invention;

[0042] Figure 2 The edge data caching flowchart provided by this invention;

[0043] Figure 3 This invention provides a model diagram for constructing individual user interests in the hotspot entity recognition method. Detailed Implementation

[0044] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0045] Please see Figure 1 , Figure 1 The diagram shows a model of an edge caching system with hotspot entity recognition, which includes the following components:

[0046] Searching users submit search requests to the edge server through their client applications. Search requests may include the current search time, search location, functional attributes of the entity, and entity status data collected by sensing devices. After obtaining the search results, users provide feedback on the search process based on their experience.

[0047] Entities: Composed of physical entities and their attached sensing devices (such as temperature sensors, humidity sensors, wearable devices, etc.). These sensing devices have communication capabilities with the outside world, are responsible for collecting the current state of the physical entities, and periodically reporting it to the edge server. Hot entities are those with high popularity, frequently requested by search users, or whose users have explicitly expressed a preference for them.

[0048] Edge servers: Edge servers are devices with strong storage and computing capabilities. They receive entity data uploaded by sensing devices, perform processing operations and cache selection, and improve search services for search users.

[0049] Please see Figure 2 , Figure 2 The diagram shows a flowchart of an edge caching method with hotspot entity recognition. This edge caching method specifically includes the following steps:

[0050] 1) Data collection: Entity data upload: Sensing devices periodically observe entity status data and selectively upload it; User search record storage: Edge servers store users' historical query content and their search rating data;

[0051] 2) Edge reasoning: The hot entity identification method designed according to the present invention calculates the popularity of an entity and determines whether the entity is a hot entity. If it is, the hot entity has the right to be cached first.

[0052] 3) Edge caching: When the total size of all hot entity information is less than the system cache capacity, all hot entity information is cached on the edge server. Non-hot entities are actively cached according to their popularity and size until the edge server cache capacity is full. When the total size of all hot entity information is greater than the system cache capacity, hot entities with higher popularity are cached first until the edge server cache capacity is full.

[0053] 4) Content replacement: The edge server periodically infers the interests of individual users and groups of users based on the user's search history, and updates the results of entity popularity and hot entity identification; at the same time, it replaces the cached entity information according to the edge caching strategy in 3).

[0054] 5) User Search: When a user sends a search request, the edge server quickly determines whether the user is searching for a hot entity and its cache location. If it is cached on the edge server, the result is returned directly to the user; otherwise, the edge server sends the user's search request to the terminal sensing device, which collects the data and returns it to the user through the edge server.

[0055] Please see Figure 3 , Figure 3 The diagram shown is a model of the hotspot entity recognition method designed in this invention. The method specifically includes the following steps:

[0056] 1) Search history storage: The edge server stores user search requests and feedback in the management area so as to analyze user interest preferences and cognition of entities by utilizing user search history, which is beneficial for subsequent edge caching strategy design;

[0057] 2) Personal interest model construction: Based on the historical request records of users searching in the same area, the user's short-term and long-term preferences are obtained according to the user's short-term and long-term search history, respectively. The two preferences are adaptively fused to obtain the final personal user preferences, construct an individual interest prediction model, and obtain the individual user's list of interests in entities.

[0058] (1) Extracting high-dimensional hidden features h of each user's short-term search history I(u) based on the LSTM model. I(u) It passes through the forgetting gate. k Input gate i k Output gate o k Cell state c k and hidden layer h k Composition, and h k h, as a high-dimensional hidden feature |I(u)| Then, by combining the attention mechanism, the degree of influence (a) of high-dimensional hidden features on the user's future interests is obtained. k To achieve weighted operations on high-dimensional hidden features to obtain a representation of user short-term preferences.

[0059] (2) Extracting long-term user preferences using the Asymmetric-SVD method Long-term user preferences This can be obtained from its long-term search history of entities. This indicates the search history for each user. The weighting factor, The feature mapping representation of the input data. v l ,b r These are the parameters learned by the model.

[0060] (3) By adaptively combining long-term and short-term user preferences through MLP, a dynamic weighting factor is output. σ(·) is the sigmoid function, W f ,b f The parameters of the concat(·) model for the fully connected network are learned, and the final user preferences are obtained. Then calculate the relationship between individual user u and unknown entity e. i level of interest

[0061] (4) The model described in (1)-(3) is optimized and trained using stochastic gradient descent (SGD), and the training loss function is: λ represents the predicted user preference value and the actual user preference value, respectively; λ||Θ||2 is the regularization term; λ is the weight factor; and Θ is the learning parameter for all models.

[0062] (5) Using the personal interest model constructed in (1)-(4), predict individual user u's relationship with unknown entity e. i level of interest And obtain the individual user's list of interests in entities.

[0063] 3) Group interest model construction: Based on the individual interest model and the known entity state, joint reasoning is performed to obtain the popularity of unknown entities.

[0064] (1) According to user u x and u y User interest list and Get user u x and u y similarity sim(u x ,u y ):

[0065]

[0066] (2) According to user u x With all other users u y Overall similarity Calculate user u x Similarity to groups

[0067] (3) Due to overall similarity Higher user preference for candidate entities can more accurately reflect entity popularity and obtain individual user interest weights.

[0068] (4) Calculate entity ei Popularity

[0069] 4) Hotspot Entity Recognition: Analyzing the popularity of entities Set a threshold ρ to distinguish between hot and non-hot entities. These are hotspot entities. This is a non-hotspot entity.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying hotspot entities, characterized in that, The method specifically includes the following steps: S1: Store user search records: The edge server stores the search history of each user within its management area. Search history ,in Indicates user exist Entities were searched at all times. ; S2: Building an Individual Interest Model: The edge server builds an individual interest prediction model based on the historical request records of users searching within the same area, and uses this model to predict each user's level of interest in unknown entities. Acquire personal users Entity Interest List ,in N This represents the total number of entities; specifically, it includes the following steps: S21: Extracting high-dimensional hidden features of each user's short-term search history based on the LSTM model. ; S22: Integrating attention mechanisms to determine the impact of high-dimensional hidden features on users' unknown entity interests Weighted fusion operations are performed on high-dimensional hidden features to form the short-term preferences of individual users. ; S23: Extract high-dimensional hidden features from users' long-term search history using asymmetric singular value decomposition and use them as long-term preferences for individual users. ; S24: Individual user interest preferences are a comprehensive reflection of long-term and short-term interest preferences. The final individual user interest preferences are obtained by adaptively combining the two preferences through a multilayer perceptron. ,in As a weighting factor, For the sigmoid function, For a fully connected network The model learns the parameters; then the individual user is calculated. For unknown entities level of interest ; S25: Optimize and train the model from steps S21-S24 using the stochastic gradient descent algorithm, with the training loss function being... ,in These are the predicted user preference values ​​and the actual user preference values, respectively. For regularization terms, As a weighting factor, Parameters are learned for all models; finally, the trained models are used to predict individual user preferences. Interest list of individuals ; S3: Constructing a group interest model: Based on the individual interest prediction model, analyze group users By performing joint reasoning on the entity interest list, we can obtain the entity interest list of group users. ,in The entity represented popularity M This represents the total number of users; specifically, it includes the following steps: S31: According to user and User interest list and Acquire users and similarity for: S32: According to user and similarity Calculate users Similarity to groups ; S33: Obtaining Individual User Interest Weights ; S34: Computational Entity Popularity , Total number of users; S4: Identify Hot Entities: Analyze Entity Popularity Set threshold To distinguish between hot and non-hot entities, if If it is a hot entity, then it is a hot entity; otherwise, it is a non-hot entity.

2. The hotspot entity recognition method according to claim 1, characterized in that, Step S21 specifically includes: the LSTM model is constructed using a forget gate. Input gate Output gate Cell state and hidden layers Composition, in which , , , , , These are the weight parameters in the LSTM network model. These are the bias parameters in the LSTM model; , The activation functions are sigmoid and tanh, respectively; each short search history of the user is considered. As LSTM input The output of the hidden layer is used as the extracted high-dimensional hidden features. .

3. The hotspot entity recognition method according to claim 2, characterized in that, Step S22 specifically includes: user exist Entities searched in real time For unknown entities Predictions have differentiated impacts; an attention mechanism is introduced to calculate the relevance weights for each interaction history. Attention score obtained Adjusting cell state in LSTM and hidden layers : ; ; for the adjusted high-dimensional hidden vector Weighted average is used to obtain user short-term preferences , ,in For model learning parameters, k This indicates traversing the set sequentially. The sequence number corresponding to the element in the middle. .

4. The hotspot entity recognition method according to claim 3, characterized in that, Step S23 specifically includes: representing user long-term preferences as , This indicates the search history for each user. The weighting factor, The feature mapping representation of the input data. These are the parameters learned by the model.

5. An edge data caching method, characterized in that, The method specifically includes the following steps: 1) Data collection: Entity data upload: Sensing devices periodically observe entity status data and selectively upload it; User search record storage: Edge servers store users' historical query content and their search rating data; 2) Edge reasoning: The hotspot entity identification method according to any one of claims 1 to 4 calculates the popularity of an entity and determines whether the entity is a hotspot entity. If it is, the hotspot entity has the right to be cached first. 3) Edge caching: When the total size of all hot entity information is less than the system cache capacity, all hot entity information is cached on the edge server. Non-hot entities are actively cached according to their popularity and size until the edge server cache capacity is full. When the total size of all hot entity information is greater than the system cache capacity, hot entities with higher popularity are cached first until the edge server cache capacity is full. 4) Content replacement: The edge server periodically infers individual user interests and group user interests based on user search history, and updates the entity popularity and hot entity identification results; at the same time, it replaces cached entity information according to the edge caching strategy in step 3); 5) User Search: When a user sends a search request, the edge server quickly determines whether the user is searching for a hot entity and its cache location. If it is cached on the edge server, the result is returned directly to the user; otherwise, the edge server sends the user's search request to the terminal sensing device, which collects the data and returns it to the user through the edge server.

Images