Popularity prediction method based on federated Bayesian learning in fog wireless access network

By employing federated Bayesian learning methods and utilizing quantization and gradient encoding techniques, the problem of high communication overhead in multi-node collaborative prediction is addressed, thereby improving the accuracy and training speed of popularity prediction and reducing errors.

CN115080888BActive Publication Date: 2024-11-12SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210771115.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-06-30
Publication Date
2024-11-12
Estimated Expiration
2042-06-30

AI Technical Summary

Technical Problem

In foggy wireless access networks, existing technologies cannot effectively solve the problems of high communication costs and high content popularity prediction errors caused by multi-node collaborative prediction.

Method used

By employing a federated Bayesian learning method, gradient information is quantized and encoded, and Hamiltonian Monte Carlo sampling is combined to achieve multi-node collaborative prediction, thereby reducing communication overhead and improving the accuracy of popularity prediction.

Benefits of technology

While reducing the communication burden, it improves the accuracy of popularity prediction and training convergence speed, and reduces prediction error.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115080888B_ABST
    Figure CN115080888B_ABST
Patent Text Reader

Abstract

The present invention discloses a popularity prediction method based on federated Bayesian learning in a fog wireless access network, comprising the following steps: 1. Constructing a content request arrival probability model based on prior knowledge of content requests; 2. Based on training data, each device updates the model gradient using random variance reduction gradient; 3. Quantizing and encoding the updated gradient and transmitting it to the server; 4. The server performs gradient aggregation and performs sampling according to the Hamiltonian Monte Carlo method; 5. Predicting the popularity of content that exists in the content library and content that does not exist in the content library. The present invention effectively utilizes the computing resources and local data sets of multiple edge nodes to efficiently train the global model. The present invention also adopts federated Bayesian learning based on quantization and encoding to achieve an effective trade-off between accuracy and communication overhead, while accelerating the convergence speed and reducing prediction errors.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of edge caching in mobile communications, and in particular relates to a popularity prediction method based on federated Bayesian learning in a fog wireless access network. Background Art

[0002] The new generation of mobile communication systems aims to provide mobile communication services with high service quality and wide coverage. It is also expected to provide new developments for various vertical industry applications, such as automobiles, public transportation, medical care, energy, agriculture, entertainment, etc. At present, user density, traffic volume and data rate will increase rapidly, so new solutions are needed to meet the needs of users and vertical industries in the next decade. Among various options, rebuilding the wireless access network is seen as a solution with development potential. Fog wireless access network is a new system architecture of 5G communication system. By utilizing the computing and storage resources on the edge nodes, some content can be temporarily stored on the edge nodes to respond to frequent access requests from user devices. This technology of storing hot content on edge nodes is called mobile edge caching. In the absence of mobile edge caching, content requests of mobile users are usually provided by remote Internet content servers provided by content providers. When users retrieve the same popular content from a remote server, the remote server must repeatedly send the same file, which may cause duplicate traffic and network congestion. Therefore, by caching popular content closer to the user, the delay in retrieving content can be greatly reduced, and repeated transmission from the content server to the network nodes connected to the cache can be avoided. However, the storage space of nodes is limited, and the popularity of content changes rapidly over time and space. Edge caching technology faces many challenges, such as what kind of content should be stored in local nodes.

[0003] Traditional caching strategies, such as the least recently used algorithm strategy, record the most recent requests for each content. When there is not enough space, it replaces the content that has not been requested for the longest time with the newly requested content. Similarly, the least frequently used strategy replaces the content that is least frequently requested. These traditional passive caching strategies and their variants are widely used in practice due to their simplicity. However, this cache replacement strategy that does not consider the popularity of the content has a general effect. The current research direction is mainly to turn to edge caching strategies based on the popularity of the content itself, but the current strategies mainly consider the prediction scheme of a single node. The very few schemes that consider multi-node collaborative prediction require high communication overhead and impose a heavy burden on the communication link. If a more accurate popularity prediction can be maintained at a lower communication cost, the prediction error can be reduced, the communication burden can be reduced, the cache hit rate can be improved, and the link load and communication delay can be greatly reduced. Summary of the invention

[0004] The present invention aims to provide a popularity prediction method based on federated Bayesian learning in a fog wireless access network to solve the technical problems of high communication costs and high errors in content popularity prediction caused by multi-node collaborative prediction.

[0005] In order to solve the above technical problems, the specific technical solution of the popularity prediction method based on federated Bayesian learning in the fog wireless access network of the present invention is as follows:

[0006] A popularity prediction method based on federated Bayesian learning in a fog wireless access network comprises the following steps:

[0007] Step 1: According to the random arrival of content requests, the number of requests for the fth content in the nth time slot satisfies the Poisson distribution And establish the corresponding probability model, where represents the number of requests for the fth content in the nth time slot, which is regarded as the content popularity, λ f (x f ) indicates that the arrival rate that obeys the Poisson distribution is the Q-dimensional content feature vector x f Function of

[0008] Step 2: Record the number of requests for the existing content in the content library from the 1st to the Nth time slot and the corresponding Q-dimensional content feature vector [x1, x2, …, x f ] T As training data, the superscript T represents the transpose of the matrix; each device updates the gradient of the model based on the local training dataset;

[0009] Step 3: quantize and encode the gradients calculated in step 2 and transmit them to the server.

[0010] Step 4: The server decodes the received results and assigns different weights for federated aggregation, and then performs sampling according to the Hamiltonian Monte Carlo method;

[0011] Step 5: Categorize the content that exists in the content library and the content that does not exist in the content library, and calculate the popularity of the content c1~cx that already exists in the content library. Indicates that the given content is observed In the case of, the expected popularity of all content in the content library in the next time slot is calculated, where, For upcoming new content that does not exist in the content library F+1 , calculate its popularity E(r F+1 [N+1]|x F+1 ), which represents the Q-dimensional content feature vector x F+1 , calculate the expected popularity of the content in the next time slot.

[0012] Furthermore, step 3 specifically includes the following steps:

[0013] Step 3.1: According to the calculated gradient result v, each component v of the vector i The result is quantized by Q i (v, s) = ||v||2·sgn(v i )ξ i (v, s) and then output, where ||v||2 represents the modulus of the vector, sgn(v i ) is used to represent the sign of the i-th component, and sgn(0) = 1, ξ i (v, s) represents the random variable used for quantization, s is the quantization level, and for each ξ i (v, s), let 0≤l i <s is a positive integer such that |v i | / ||v||2∈[l i / s,l i +1 / s],ξ i The definition of (v, s) is as follows:

[0014]

[0015] In the formula Where p represents the probability of taking a value;

[0016] Step 3.2: For each gradient, only the vector modulus ||v||2 needs to be transmitted once; for each component, the quantization result Q i (v, s), and its corresponding quantization level s·ξ i (v, s) is recursively encoded by ELIAS and then transmitted to the server; in addition, Q i The sign of (v, s).

[0017] Furthermore, step 4 specifically includes the following steps:

[0018] Step 4.1, decode the received encoding result to obtain a gradient vector;

[0019] Step 4.2: According to the data set size of each federated learning participant, the corresponding aggregation weight is assigned for aggregation to obtain the global gradient, which aggregates the gradient information of each federated learning participant;

[0020] Step 4.3: Perform Hamiltonian Monte Carlo sampling based on the global gradient, iterate the sampling samples, and obtain S samples with stable distribution

[0021] The popularity prediction method based on federated Bayesian learning in the fog wireless access network of the present invention has the following advantages:

[0022] 1. The federated Bayesian learning method proposed in the present invention uses the contents of steps 1, 2, and 3 to effectively utilize the computing resources and local data sets of multiple edge nodes to efficiently train the global model. In the current environment, the coverage considered by the problem will be much larger than the coverage of a single node, and the number of associated users will be greatly improved.

[0023] 2. Since multi-node collaborative prediction requires mutual communication to exchange training results, it will bring about large communication overhead and increase the burden on the communication link. Therefore, this method uses the contents of steps 3, 4, and 5, and adopts federated Bayesian learning based on quantization and coding, which can achieve an effective trade-off between accuracy and communication overhead, and at the same time can accelerate the convergence speed and reduce prediction errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of a popularity prediction method based on federated Bayesian learning in a fog wireless access network of the present invention;

[0025] Figure 2 A schematic diagram showing the relationship between the prediction error and the number of observation time slots of the present invention;

[0026] Figure 3 This is a schematic diagram of the training convergence speed comparison of the present invention;

[0027] Figure 4 It is a schematic diagram for comparing the number of bits consumed in communication of the present invention. DETAILED DESCRIPTION

[0028] In order to better understand the purpose, structure and function of the present invention, the popularity prediction method based on federated Bayesian learning in a fog wireless access network of the present invention is further described in detail below with reference to the accompanying drawings.

[0029] A popularity prediction method based on federated Bayesian learning in fog wireless access networks, such as Figure 1 The following steps are shown:

[0030] Step 1: According to the random arrival of content requests, the number of requests for the fth content in the nth time slot satisfies the Poisson distribution And establish the corresponding probability model, where represents the number of requests for the fth content in the nth time slot, which can be roughly regarded as the content popularity, λ f (x f ) indicates that the arrival rate that obeys the Poisson distribution is the Q-dimensional content feature vector x f Function of

[0031] Step 2: Record the number of requests for each historical time slot for the existing content in the content library The superscript T represents the transpose of the matrix and the corresponding Q-dimensional content feature vector [x1, x2, ..., x f ] T As training data, each device updates the gradient of the model based on the local training dataset;

[0032] Step 3: quantize and encode the gradient calculated in step 2 and transmit it to the server.

[0033] Step 3.1: According to the calculated gradient result v, each component v of the vector i The result is quantized by Q i (v, s) = ||v||2·sgn(v i )ξ i (v, s) and then output, where ||v||2 represents the modulus of the vector, sgn(v i ) is used to represent the sign of the i-th component, and sgn(0) = 1, ξ i (v, s) represents the random variable used for quantization. i (v, s), let 0≤l i <s,l i is a positive integer such that |v i | / ||v||2∈[l i / s,l i +1 / s],ξ i The definition of (v, s) is as follows:

[0034]

[0035] In the formula Where p represents the probability of taking a value;

[0036] Step 3.2: For each gradient, only the vector modulus ||v||2 needs to be transmitted once; for each component, the quantization result Q i (v, s), and its corresponding quantization level s·ξ i (v, s) is recursively encoded by ELIAS and then transmitted to the server; in addition, Q i The sign of (v, s).

[0037] Step 4: The server decodes the received results and assigns different weights for federated aggregation, and then performs sampling according to the Hamiltonian Monte Carlo method;

[0038] Step 4.1, decode the received encoding result to obtain a gradient vector;

[0039] Step 4.2: According to the data set size of each federated learning participant, the corresponding aggregation weight is assigned for aggregation to obtain the global gradient, which aggregates the gradient information of each federated learning participant;

[0040] Step 4.3: Perform Hamiltonian Monte Carlo sampling based on the global gradient, iterate the sampling samples, and obtain S samples with stable distribution

[0041] Step 5: Classify the content to be predicted based on whether it exists in the content library or not. For the content that already exists in the content library, calculate its popularity in the next time slot. Indicates that the given content is observed In the case of, the expected popularity of all content in the content library in the next time slot is calculated, where, For upcoming new content that does not exist in the content library F+1 , calculate its popularity E(r F+1 [N+1]|x F+1 ), which represents the Q-dimensional content feature vector x F+1 , calculate the expected popularity of the content in the next time slot.

[0042] Figure 2 It is a schematic diagram of the relationship between the prediction error and the number of observation time slots of the present invention, which shows that the present invention can obtain a lower prediction error under the same observation time slot.

[0043] Figure 3 This is a schematic diagram for comparing the training convergence speed of the present invention, which shows that the convergence speed of the present invention is better than that of the traditional solution.

[0044] Figure 4 This is a schematic diagram comparing the number of communication bits consumed in the present invention, which shows that compared with the traditional solution, the present invention only requires fewer communication bits to obtain the same prediction accuracy.

[0045] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A popularity prediction method based on federated Bayesian learning in fog wireless access network, characterized in that: The following steps are involved: Step 1: According to the random arrival of content requests, the number of requests for the fth content in the nth time slot satisfies the Poisson distribution And establish the corresponding probability model, where represents the number of requests for the fth content in the nth time slot, which is regarded as the content popularity, λ f (x f ) indicates that the arrival rate that obeys the Poisson distribution is the Q-dimensional content feature vector x f Function of Step 2: Record the number of requests for existing content in the content library from the 1st to the Nth time slots and the corresponding Q-dimensional content feature vector [x1, x2, …, x f ] T As training data, the superscript T represents the transpose of the matrix; each device updates the gradient of the model based on the local training dataset; Step 3: quantize and encode the gradients calculated in step 2 and transmit them to the server. Step 4: The server decodes the received results and assigns different weights for federated aggregation, and then performs sampling according to the Hamiltonian Monte Carlo method; Step 5: Categorize the content that exists in the content library and the content that does not exist in the content library. F , calculate its popularity Indicates that the given content is observed In the case of, calculate the expected popularity of all content in the content library in the next time slot, where, For upcoming new content that does not exist in the content library F+1 , calculate its popularity E(r F+1 [N+1]|x F+1 ), which represents the Q-dimensional content feature vector x F+1 , calculate the expected popularity of the content in the next time slot; The step 3 specifically comprises the following steps: Step 3.

1. According to the calculated gradient result v, each component v of the vector i The result passes through the quantizer Q i (v, s) = ||v||2 · sgn(v i )ξ i (v, s) and then outputs, where ||v||2 represents the norm of the vector, sgn(v i ) is used to represent the sign of the i-th component, it is agreed that sgn(0) = 1, ξ i (v, s) represents the random variable for quantization, s is the quantization level, for each ξ i (v, s), let 0 ≤ l i < s be a positive integer such that |v i | / ||v||2 ∈ [l i / s, l i +1 / s], the definition of ξ i (v, s) is as follows: In the formula Where p represents the probability of taking a value; Step 3.2: For each gradient, only the vector modulus ||v||2 needs to be transmitted once; for each component, the quantization result Q i (v, s), and its corresponding quantization level s·ξ i (v, s) is recursively encoded by ELIAS and then transmitted to the server; in addition, Q i The sign of (v, s).

2. The popularity prediction method based on federated Bayesian learning in fog wireless access network according to claim 1, characterized in that: The step 4 specifically comprises the following steps: Step 4.1, decode the received encoding result to obtain a gradient vector; Step 4.2: According to the data set size of each federated learning participant, the corresponding aggregation weight is assigned for aggregation to obtain the global gradient, which aggregates the gradient information of each federated learning participant; Step 4.3: Perform Hamiltonian Monte Carlo sampling based on the global gradient, iterate the sampling samples, and obtain S samples with stable distribution