A Content Caching and Recommendation Method Based on Federated Learning in a Fog Computing Network

Abstract

The present invention relates to a content caching and recommendation method based on federated learning in a fog computing network, belonging to the field of communication technologies. In this network, each fog node FN will cache content with high popularity. At the same time, Internet of Things devices ID, as clients of federated learning, can locally cache some content and use local data to train model parameters, avoiding the direct upload of private information of the IDs. Due to the limited storage resources at the ID side and the inherent user demand patterns, the content caching gain is limited. The IDs can obtain content through D2D cooperative links or through the FN or the cloud. To improve the caching utility at the ID side, the KNN (K-Nearest Neighbors) algorithm is used to find neighbor IDs and recommend cached content to the target ID, and the target ID caches the content according to the content score. To improve the cache hit rate, the FN will establish a personalized content recommendation list, track user needs through active content recommendation, and reduce the content acquisition delay.

Description

Technical Field

[0001] The present invention belongs to the field of communication technologies, and relates to a content caching and recommendation method based on federated learning in a fog computing network. Background Art

[0002] With the development of the Internet of Things and communication technologies, the number of next-generation Internet of Things devices has increased explosively, and the mobile traffic has also increased several times. Along with the emergence of more and more new applications, users have a large number of choices and have higher and higher requirements for content request latency. However, the current surge in mobile traffic is causing the time for users to obtain content to continue to increase, thus bringing huge pressure to the local base stations and the backhaul links of the Internet. Therefore, the research on edge caching has become one of the hottest research topics in the field of wireless communication. Based on content popularity, popular content can be cached in fog nodes or user local areas closer to users, which can effectively reduce network link congestion and request latency, thereby improving the quality of service QoS of users.

[0003] In recent years, fog computing (FC) is a new type of edge computing network framework. By pushing computing and storage functions to the network edge, closer to users, it realizes the extension of cloud computing to the network edge to support the growing demand for multimedia services. Compared with cloud computing, distributed edge caching among fog nodes (FNs) can effectively reduce network traffic and service latency because it places popular content in FNs instead of remote clouds. However, due to the limited coverage and caching resources of FNs, a cooperative caching scheme is needed to determine which content is popular and how to cache it. For example, First In First Out (FIFO), Least Recently Used (LRU), and end-to-end Device-to-Device (D2D) cooperative caching. These methods usually collect and analyze data from Internet of Things devices (IDs) on a central server to predict the popularity of cached content, which will consume a large amount of communication bandwidth and cause serious privacy problems.

[0004] To solve the above problems, federated learning (FL) is considered an effective method. FL is a decentralized framework that uses training data distributed on Internet of Things devices to collaboratively learn a model to improve communication efficiency. Basically, the IDs update local model parameters through local data, and the FN side aggregates the uploaded updated model parameters to complete the update of global model parameters to learn a shared training model. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a content caching and recommendation method based on federated learning in a fog computing network.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] A content caching and recommendation method based on federated learning in a fog computing network, the method includes the following steps:

[0008] S1: D2D Collaboration Model Based on IDs

[0009] S2: Local Content Caching Model Based on Deep Neural Network

[0010] S3: Active Caching Algorithm Based on Federated Learning

[0011] S4: Personalized Content Recommendation Algorithm Based on IDs

[0012] In step S1, establishing D2D communication connections between IDs can reduce the traffic burden on the FN and the cloud. The D2D collaboration model mainly considers two factors, namely the physical link quality and social intensity between IDs.

[0013] (1) The physical link quality is considered from the following two aspects: First, the D2D contact time needs to exceed the content transmission time; Second, the IDs establishing D2D connections need to meet their transmission rate requirements.

[0014] The D2D contact time T com represents the time from the establishment of the connection to the disconnection of the connection between ID u and v, and is assumed to follow an exponential distribution. R u，v represents the transmission rate of ID u and v through the D2D link, and is expressed as:

[0015]

[0016] where B u，v represents the transmission bandwidth allocated by ID u to ID v, p v represents the transmission power consumption of ID v, H u，v represents the channel gain between ID u and v, H u，i represents the channel gain between ID u and other ID i, δ 2 represents the power of Gaussian white noise.

[0017] T u，v，c represents the transmission time of content c by ID u and v through the D2D link, and is expressed as:

[0018]

[0019] where s c represents the size of content c. To ensure successful content acquisition, the contact time needs to exceed the content transmission time, and its probability is expressed as:

[0020]

[0021] In addition, considering that the transmission rate requirements for different contents in IDs are different, the probability of meeting the transmission rate requirements is represented by the Logistic function:

[0022]

[0023] Among them, R c，th represents the transmission rate requirement of content c, and τ represents the slope parameter of this logistic function.

[0024] (2) To establish a stable D2D connection, the social strength needs to consider the interest similarity, contact strength, and social strength among IDs simultaneously.

[0025] The preference of ID u for content c is represented as q u，c , |q u，c -q v，c |≤Δ is considered that ID u and v have the same preference for content c, where Δ takes a small positive number. The binary variable φ u，v，c is introduced to represent the interest similarity between ID u and v for content c:

[0026]

[0027] Then the interest similarity between ID u and v is represented as:

[0028]

[0029] The contact history strength between IDs is also an important factor in measuring social strength. The contact strength of IDs at time t is represented as ρ and z are control parameters. Therefore, at time t now , the normalized contact strength between ID u and v is represented as:

[0030]

[0031] Among them, k u，v , represents the number of contacts between ID u and v, represents the contact time between ID u and v. In addition, ξ u，v ∈[0, 1] represents the social relationship between ID u and v. In summary, the social strength between ID u and v is:

[0032] E u，v =φ u，v ι 1 +I u，v ι 2 +ξ u，v ι 3

[0033] Among them, ι 1 , ι 2 , ι3 ∈ [0, 1], ι 1 + ι 2 + ι 3 = 1 is an adjustable hyperparameter, representing the social intensity weights of IDs u and v.

[0034] Furthermore, in step S2, the DNN training model of IDs consists of an input layer, a hidden layer, and an output layer. A large number of historical request data records in IDs are used as training samples, denoted as where x c = [x c，1 , x c，2 ,..., x c，N T , representing the N-dimensional features of content c; Denotes the score (preference degree) of ID u for content c in the t-th round of communication. The Sigmoid function is used as the activation function of the hidden layer, that is where k represents the k-th layer of the DNN model; the expected output of the DNN model is the preference prediction value of ID u for content c, that is

[0035] During the local training process of IDs, in the ι-th round of local training in the t-th round of global aggregation, the update of the DNN model parameters is expressed as:

[0036]

[0037] where η represents the learning rate, respectively represent the exponential moving average and square of the local model parameter gradient:

[0038]

[0039]

[0040] where, b 1 , b 2 ∈ [0, 1) represents the exponentially decaying estimate at the current moment, is the bias-corrected estimate.

[0041] Then the model cross-entropy loss function is expressed as:

[0042]

[0043] The goal of DNN model training is to minimize the loss function, that is ​Based on the trained DNN model, through the method of KNN retrieval, the content recommended by neighbor IDs is used as the input of the DNN model for the target ID, and the content with a large content preference is actively cached under the condition of cache capacity limitation.

[0044] Furthermore, in step S3, while the IDs are uploaded the number of requests for content by ID u in the time period [t - T0 + 1, t] is also uploaded. The activity level of ID u at the FN side is expressed as:

[0045]

[0046] The FN side performs model aggregation, which is expressed as: When the FL training reaches the global accuracy, the FN predicts the content popularity according to the aggregated global model and caches the content with a higher popularity under the limitation of the cache capacity.

[0047] Furthermore, in step S4, the update of the recommendation list of ID u is related to the current cache status and the candidate recommendation list. A personalized user recommendation algorithm based on the simulated annealing algorithm (SA) is proposed to avoid local optima and achieve global optima. The state transition probability in SA is represented by the Metropolis algorithm, and the optimal solution of the combinatorial optimization problem is obtained by repeatedly executing the Metropolis algorithm. Based on the Metropolis criterion, the recommendation list R u of ID u is updated to R u '. The state transition probability is expressed as:

[0048]

[0049] where B c is a constant, and ΔT u represents the total delay difference in obtaining the recommended content for the two combinations of R u and R u ', which is expressed as:

[0050] ΔT u = T u (R u ') - T u (R u )(ΔT u > 0)

[0051] represents the minimum total delay in obtaining the currently recommended content for ID u, which is expressed as:

[0052]

[0053] As the number of iterations increases, decreases. When Tu (R u ′) > T u (R u ), ΔT u The larger, h u (R u , R u ′) is smaller, that is, the probability of updating the recommendation list is smaller. When the number of iterations reaches a sufficiently large value, it can be considered that the probability of updating the recommendation list is almost zero at this time, and the optimal recommendation list is obtained.

[0054] The steps for updating the recommendation list are as follows:

[0055] Step 1: Determine the candidate items M of the recommendation list according to the estimated preference of ID u and the real-time request pattern of the user u , and set the maximum number of iterations T max , where

[0056] Step 2: When the number of iterations t < T max , randomly select content i′ ∈ Mu\R u , and replace each content in the recommendation list R u , then R u ′ = R u \{i} ∪ {i′};

[0057] Step 3: Update the recommendation list to R u (R u , R u ′) with a probability of h u ′;

[0058] Step 4: Update R u = R u ′, and repeat Step 2.

[0059] The beneficial effects of the present invention are as follows: In the face of the limitation of cache resources, the FN side and the ID side can cache content respectively according to the content popularity and the preference of the ID for the content. At the same time, in order to increase the request probability of the ID for low-latency content, a personalized content recommendation scheme based on IDs is designed to recommend attractive content according to the real-time behavior of the user, improve user satisfaction, reduce the content acquisition latency, and increase the cache hit rate.

[0060] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent description, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0062] Figure 1 It is a content caching model based on FL in a fog computing network.

[0063] Figure 2 It is a flow chart of a collaborative content caching and recommendation algorithm based on FL. Specific embodiments

[0064] The following specific examples illustrate the embodiments 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. Various details in this specification can also 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 drawings provided in the following embodiments only illustrate the basic concept of the present invention schematically. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0065] Among them, the accompanying drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, and do not represent the dimensions of actual products; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0066] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and should not be construed as a limitation to the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0067] Figure 1The network topology diagram is a three-layer network consisting of cloud-fog-IoT. The FN is equipped with a certain amount of caching resources and is connected to the cloud through the Mobile Network Operator (MNO) center. The IDs are equipped with an AI chipset that supports offline learning in intelligent wireless communication scenarios and have a certain amount of computing resources and caching resources. The set of FNs is represented as M = {1, 2,..., m,..., M}, and the caching space is represented as The set of IDs is represented as U = {1, 2,..., m..., U}, and the caching space is represented as where The content provided by the cloud is represented as C = {1, 2,..., c,..., C}, and the size of content c is represented as s c . In the actual scenario, different types of content have different sizes and corresponding QoS requirements. The IDs first obtain content locally. If the content acquisition fails, they obtain information from neighboring IDs through D2D links, or obtain information from the FN and the cloud.

[0068] 1. Caching Model

[0069] The FN will regularly predict the content popularity to update the proactive caching scheme according to the caching capacity. The content popularity depends on the personal preferences and activity levels of the IDs and is represented as p = {p 1 , p 2 ,..., p c ,..., p c}, where p c represents the probability that the IDs within the FN coverage request content c. The preference of ID u for content c is represented as q u，c ∈[0, 1], and satisfies Considering that the IDs with a high activity level request content more frequently, the activity level of the IDs is represented as α = {α 1 , α 2 ,..., α u ,..., α u}, where α u ∈[0, 1] represents the probability that ID u requests content and satisfies

[0070] In the actual scenario, as the IDs continuously initiate content requests, the content popularity will change dynamically, and the cached content will be updated regularly. Since the FN caching space is limited, the content with a high popularity is preferentially cached. The caching decision at the FN side is represented as ψ m，c ∈{0, 1}, indicating whether FN m caches content c. When ID u requests content c from FN m and ψ m，c = 1, then the FN m cache hits; when ψm，c = 0, FN m will compare the popularity of the content c with that of other cached content and consider whether to cache the content c so that the IDs can obtain the content next time.

[0071] 2. Recommendation Model

[0072] 2.1 Recommendation Model

[0073] Based on the traditional content collaborative filtering recommendation algorithm (CF), low-latency and attractive content is recommended to the IDs. The content recommended by IDu is related to its own content preference, the latency of obtaining content, and real-time behavior. Usually, the cosine similarity between ID u and content c is used to represent the estimated preference degree of ID u for content c, which is related to the content feature vector and the user feature vector, and is expressed as:

[0074]

[0075] The content is represented by an N-dimensional feature, s u，c The larger it is, the greater the preference degree of ID u for content c and the greater the request probability. e u，n ∈ [0, 1], n ∈ {1, 2,..., N} is the user feature vector, representing the preference degree of ID u for feature n; h c，n ∈ [0, 1], n ∈ {1, 2,..., N} is the content feature vector, representing the correlation degree between content c and feature n. Represents the normalized preference degree of ID u for content c. The patience threshold k of the user is introduced u , when p u，c > k u it becomes a recommendation candidate item. From this, the initial candidate recommendation list can be obtained, and then the updated candidate recommendation list M is obtained according to the real-time behavior of ID u u . Since the screen size of mobile devices is limited, the recommended content is a content combination that minimizes latency in the current cache state, thereby obtaining the final recommendation list R of ID u u .

[0076] 2.2 Content Request Model

[0077] Based on the recommendation model, content is recommended to the IDs. Generally, it will increase the request probability of the IDs for the recommended content. The request probability of IDu for the recommended content c is affected by content preference and the order of the recommendation list. The Zipf distribution is used to describe the influence of the recommendation list order on ID u's request for content c, and is expressed as:

[0078]

[0079] where β u is the allocation coefficient of the u-th user, R is the number of contents in the recommendation list, Zu，c and L uci are both binary variables. Z u，c = 1 indicates that content c is in the recommendation list of ID u, while L uci = 1 indicates that content c is in the i-th position in the recommendation list of ID u. All elements Z u，c constitute the recommendation strategy matrix Z U×C .

[0080] The influence of content preference in the recommendation list on the content c requested by ID u is expressed as:

[0081]

[0082] Then the request probability of ID u for content c in the recommendation list is expressed as:

[0083] p rec (u, c) = p ListRec (u, c) · p PrefRec (u, c)

[0084] The user can accept or reject the recommended content. Use γ u to represent the probability of accepting the recommendation list. γ u is related to the historical probability p Arec (u) of ID u accepting the recommendation, the probability p Rrec (u) of successfully requesting the recommended content, and the deviation of the recommended content preference. Among them, the deviation of the recommended content preference of ID u is also called preference distortion D u :

[0085]

[0086]

[0087]

[0088] where ω uci ∈ {0, 1}, ω uci = 1 indicates that the c-th content is the i-th content item in the initial preference list of ID u, arranged in descending order. Then:

[0089] γ u = f 1 p Arec (u, c) + f 2 p Rrec (u, c) + f 3 (1 - D u )

[0090] where f 1 , f 2 , f 3 ∈ [0, 1], f1 +f 2 +f 3 = 1 is an adjustable hyperparameter representing γ u The intensity weight of the relevant parameters. Then the request probability of user u for content f is:

[0091] p req (u, c) = γ u ·p rec (u, c) + (1 - γ u )p u，c

[0092] 3. Communication Model

[0093] The cache placement decision of the content is represented as ψ c = {Ψ u，c , Ψ m，c , Ψ cloud，c}. Among them, Ψ u，c， Ψ m，c , Ψ cloud，c ∈ {0, 1}, indicating whether content c is cached at the ID side, FN side, or cloud side respectively, is the maximum cache capacity of ID u , and similarly is the maximum cache capacity of FN m ; The cache delivery decision of content c is represented as Θ c = {θ u，0，c , θ u，v，c , θ n，m，c， θ u，cloud，c}. Among them, θ u，0，c , θ u，v，c , θ u，m，c , θ u，cloud，c ∈ {0, 1}, indicating whether ID u obtains content c from local, D2D link, FN m or cloud respectively. The premise of content delivery is that the device has cached the corresponding content and satisfies θ u，0，c + θ u，v，c + θ u，m，c + θ u，cloud，c = 1.

[0094] The user first considers whether the requested content is cached locally. The latency for ID u to obtain content locally is negligible; when θ u，0，c = 0, ID u considers the latency to obtain content c from ID v, which is represented as:

[0095]

[0096]

[0097] o u，vDenote the connection parameter of IDs u and v, Γ 1 is the social intensity threshold, Γ 2 , Γ 3 is the physical link quality threshold.

[0098] When θ u，0，c = 0, θ u，v，c = 0, the transmission rate and delay for ID u to obtain content c from FN m are respectively expressed as:

[0099]

[0100]

[0101] Among them, B u，m , represents the transmission bandwidth allocated by FN m to ID u, P m represents the transmission power of FN m, H u，v represents the channel gain between FNm and ID u.

[0102] When θ u，0，c = 0, θ u，v，c = 0, θ u，m，c = 0, ID u can only obtain the requested content from the cloud. Here, it is assumed that the time for all IDs to obtain content from the cloud is equal, and T cloud >> T u，v，c , T u，m，c .

[0103] 4 Establish an optimized caching scheme:

[0104]

[0105] S.t.C1: E u，v ≥ Γ 1

[0106] C2: pr(T com > T u，v，c )≥ Γ 2

[0107] C3: Pr Qos ≥ Γ 3

[0108] C4:

[0109] C5:

[0110] C6: o u，v ∈ {0, 1}

[0111] C7:

[0112] C8: Ψ u，c， Ψ m，c , Ψ cloud，c ∈ {0, 1}

[0113] C9: θ u，0，c , θ u，v，c , θ u，m，c , θ u，cloud，c ∈ {0, 1}

[0114] C10: θ u，0，c + θ u，v，c + θ u，m，c + θ u，cloud，c = 1

[0115] Where Ψ represents the cache placement decision, and Θ represents the cache delivery decision. The constraints C1, C2, and C3 represent the constraints that need to be satisfied to establish a stable link in the D2D cooperation model. The constraints C4 and C5 represent the cache capacity constraints of the ID side and the FN side respectively. The constraints C6 and C7 represent the D2D connection constraints, that is, the IDs can obtain content from at most one associated ID through D2D connection. The constraints C8, C9, and C10 represent the cache placement and delivery decision constraints, that is, the content is indivisible and can only be obtained through one of the paths of its own cache, D2D user cache, FN cache, or cloud.

[0116] To reduce the latency of information acquisition by IDs and avoid information leakage and security risks caused by a large amount of information upload, an FL distributed framework is introduced to learn global parameters on the server side. The IDs share the local model parameters trained locally with the server without uploading local data.

[0117] Figure 2 For the content caching and recommendation scheme based on federated learning in the fog computing network, the specific steps are as follows:

[0118] Step 201: Algorithm initialization.

[0119] Step 202: The FN sets the information monitoring period and monitors the set of IDs within its coverage in discrete time periods. The IDs download the global model and the user personalized recommendation list from the FN.

[0120] Step 203: The IDs make content requests, and the ID request model is updated as the recommendation list is updated.

[0121] Step 204: Establish the D2D link of the IDs. When o u，v = 1, establish the connection between ID u and v. If multiple IDs have the required content at the same time, obtain the content from the ID that is closer.

[0122] Step 205: According to the historical request information of the IDs, obtain the local model parameter w through an offline user preference learning algorithm.

[0123] Step 206: Retrieve neighbor users based on KNN, and the neighbor users perform cached content recommendation.

[0124] Step 207: Use the content recommended by the neighbor users as the input for the target ID, and the ID side caches the content according to the content score.

[0125] Step 208: If the local accuracy is reached at the ID side, stop the local model training; otherwise, repeat Steps 203 to 207.

[0126] Step 209: Upload the important gradients after clustering quantization at the ID side to the FN based on the k-means-based gradient compression algorithm, and upload the content request count and feature preferences of the ID s .

[0127] Step 210: The FN side calculates the activity of the uploaded IDs.

[0128] Step 211: The FN aggregates the local model parameters of the uploaded IDs.

[0129] Step 212: With the online popularity prediction algorithm, the FN side caches the content according to the content popularity.

[0130] Step 213: Based on the user's feature preferences and the content's feature preferences, according to the real-time request behavior, the FN side establishes a personalized recommendation list for the IDs.

[0131] Step 214: When the global accuracy is reached at the FN side, this is the final caching decision. If the global accuracy is not reached, repeat Steps 202 to 213.

[0132] Step 215: The algorithm ends and outputs the optimization result.

[0133] 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 them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A content caching and recommendation method based on federated learning in a fog computing network, characterized in that: This method includes the following steps: S1: A D2D cooperation model based on IDs; S2: A local content caching model based on a deep neural network; S3: An active caching algorithm based on federated learning; S4: A personalized content recommendation algorithm based on IDs; In the said S1, a stable D2D connection is established according to the physical link quality and social intensity between IDs; the interest similarity, contact intensity and social trust degree between IDs are considered in the social intensity, and the D2D contact time and the QoS requirements of IDs are considered in the physical link quality; when the social intensity is greater than the social intensity threshold and the physical link quality is greater than the physical link quality threshold, a D2D connection is established and content is shared; In the said S2, a deep neural network DNN model is established at the ID side, and the local model parameters W are trained using historical data and content N-dimensional features while obtaining content preferences, and M neighbor IDs of the target ID are found based on the KNN proximity algorithm for caching content recommendation. Considering the occupation and age of IDs, the recommended content is used as the DNN input of the target ID, and content caching at the ID side is performed according to the content score of the target ID; In the said S3, federated learning FL is a distributed framework that updates model parameters locally at users and aggregates the global model at the server side, without uploading local data, avoiding user privacy leakage and reducing communication bandwidth requirements at the same time; the model parameters uploaded by IDs are aggregated at the FN side, and the global model is aggregated based on the activity of users. The most popular content is cached based on the online popularity prediction algorithm; to further reduce communication overhead, a compression algorithm based on K-means is proposed to compress the uploaded model parameters; In the said S3, the gradient compression algorithm of K-means consists of two steps: first, the uploaded model gradients are divided according to the gradient value size. When the gradient value is greater than 0, it is an important gradient, and when the gradient value is approximately equal to 0, it is a secondary gradient; second, the important gradients are clustered, and the centroid value of the jth cluster is obtained using the average value in the same gradient set to approximate its gradient, and the ID only uploads the centroid value to reduce communication traffic; In the said S4, at the FN side, the patience threshold of the ID is introduced. When the preference of the ID for the content is greater than the patience threshold, it becomes a candidate for the recommendation list. Then, the request content difference degree of the ID within the t time slot is considered, that is, the variance of the preferences of each request content. If it is greater than the limit value, it means that the user is more willing to request dissimilar content, and the candidate list will remove the content similar to the request content in the t time slot; if it is less than a limit value, the candidate list will remove the content dissimilar to the request content in the t time slot from the primary candidates, and then obtain a personalized recommendation list based on the optimal recommendation algorithm; at the same time, at the ID side, the probability of the ID accepting the recommendation list is determined by the probability of previously accepting the recommendation, the probability of successfully requesting the recommended content, and the deviation of the recommended user preferences.

Images