A Fairness-Aware Edge Data Caching Method Based on Recommendation Systems

Abstract

This invention discloses a fairness-aware edge data caching method based on a recommender system, comprising the following steps: S1, obtaining historical request data of users based on the recommender system, extracting data features from the historical request data, and calculating the similarity between data features; S2, modeling user requests using a Hawkes process; predicting future user requests using the Hawkes process model; S3, quantitatively evaluating the overall user experience quality based on the obtained data feature information, and optimizing the caching strategy based on the predicted user requests using a heuristic algorithm; S4, providing a quantitative index of fairness coefficient based on the obtained data feature information and user experience quality, and optimizing the caching strategy in S3 based on this quantitative index. This invention can improve the overall system performance while ensuring service balance among different users, thereby providing a more comprehensive guarantee for user experience in edge computing environments.

Description

Technical Field

[0001] This invention belongs to the field of data caching technology, and in particular relates to a fair-aware edge data caching method based on a recommendation system. Background Technology

[0002] In recent years, with the widespread adoption of mobile devices and their enhanced computing and network capabilities, a plethora of resource-intensive applications have emerged, such as high-quality video streaming, virtual reality (VR), augmented reality (AR), and interactive games. This surge in applications has not only significantly increased mobile data traffic but also placed higher demands on the low latency and high stability of services. Statistics show that global mobile data traffic continues to grow at an average annual rate of approximately 40%, posing significant challenges to network bandwidth, computing resources, and data transmission. Traditional cloud computing's centralized processing model, due to geographical limitations, cannot adequately meet users' needs for low-latency services.

[0003] To address the high latency issues inherent in cloud computing, Mobile Edge Computing (MEC) has been proposed as a novel distributed computing architecture. In MEC, edge servers are deployed at the network edge, close to users. By caching data or performing computational tasks near users, data transmission latency can be significantly reduced, thereby improving user experience. Specifically, edge servers in an MEC environment possess caching capabilities, proactively caching data that users may request to enable rapid responses when requests are made, reducing data retrieval time and improving service real-time performance and user satisfaction.

[0004] For example, Chinese patent document CN112105035A discloses a method and apparatus for deploying mobile edge computing servers, which can ensure that each base station location can be covered by mobile edge computing servers at other base station locations, thus ensuring the QOE performance of the edge network.

[0005] However, due to the limited cache space and geographical coverage of edge servers, optimizing edge data caching strategies has become a major research challenge. Current edge caching strategies mainly focus on improving cache hit rate, reducing bandwidth requirements, and optimizing resource utilization, such as cache allocation based on content popularity prediction and priority cache allocation based on user demand frequency.

[0006] Nevertheless, most existing research focuses primarily on reducing service latency, improving cache hit rates, and reducing bandwidth consumption, often neglecting the differences in Quality of Experience (QoE) among users and the fairness of caching strategies. In resource-constrained environments, some users' requests may be prioritized, while others may suffer a poor service experience due to insufficient resources, leading to QoE imbalance. In service-intensive application scenarios, such as video streaming and AR / VR applications, QoE imbalance directly impacts overall user experience and satisfaction.

[0007] Therefore, how to improve cache hit rate while taking into account the service experience of different users has become a key issue in edge caching technology. Summary of the Invention

[0008] This invention provides a fairness-aware edge data caching method based on a recommender system to address the problems of insufficient utilization of user features and lack of fairness in caching strategies during edge caching.

[0009] A fairness-aware edge data caching method based on a recommender system includes the following steps:

[0010] S1. Based on the historical request data of users obtained from the recommendation system, extract data features from the historical request data and calculate the similarity between the data features;

[0011] S2. User requests are modeled using a Hawkes process; during the modeling process, an exponential decay function is selected as the kernel function, exhibiting monotonically decreasing characteristics, and the parameters in the Hawkes process model are estimated using the maximum likelihood estimation method; the Hawkes process model is used to predict future user requests.

[0012] S3. Based on the data feature information obtained in S1, the overall user experience quality is quantitatively evaluated. Then, based on the prediction results of user requests obtained in S2, a heuristic algorithm is used to optimize the caching strategy: prioritize the cache pairs that maximize the overall experience quality until the cache space is full, thereby finding an approximate optimal solution that maximizes the overall user experience quality.

[0013] S4. Based on the data feature information obtained in S1 and the user experience quality calculated in S3, a quantitative index of fairness coefficient is given, and the caching strategy in S3 is optimized and adjusted based on this quantitative index, so as to ensure the high quality of the overall user experience while taking into account the fairness of the experience among users.

[0014] The specific process of step S1 is as follows:

[0015] The frequency of user requests is used as a score for the data, and a user-data score matrix M is constructed based on historical requests;

[0016] The user-data rating matrix is ​​decomposed using a nonnegative matrix factorization algorithm to obtain the user feature matrix P and the data feature matrix Q.

[0017] An iterative update algorithm is used to optimize the user feature matrix P and the data feature matrix Q; the similarity between the data is calculated.

[0018] The goal of the nonnegative matrix factorization algorithm is to find matrices P and Q such that the difference between the product of P and Q and M is minimized. The formula is:

[0019]

[0020] In the formula, M represents the difference between the products of matrix M and matrices P and Q. ij This indicates the number of times user i requests data j.

[0021] An iterative update algorithm is used to optimize the user feature matrix P and the data feature matrix Q. Specifically, loss calculation and gradient descent are employed for iterative updates, as follows:

[0022]

[0023] in, This represents the value in the i-th row and j-th column of matrix P in the N-th iteration. This represents the value in the i-th row and j-th column of matrix Q in the N-th cycle.

[0024] The similarity between data is calculated by using the cosine similarity between the feature matrices of the data, as shown in the following formula:

[0025]

[0026] In the formula, Sim(d m ,d n ) represents d m ,d n The similarity between them, cos(Q) m Q n Let |U| be the cosine similarity of the vectorized representations of the two data points. m ∩U n | represents the intersection of the user sets requesting these two data sets, and Θ is the confidence threshold to avoid high similarity due to too few records.

[0027] In step S2, during the Hawkes process, the probability of an event occurring at a given time is influenced by the occurrence of similar events in the past. The intensity function of time slot t is expressed as:

[0028]

[0029] Among them, H tn The set of timestamps represents historical events, β is the background intensity, and φ(t-τ) is the kernel function that captures the impact of past events on the arrival rate of similar events.

[0030] The parameters in the Hawkes process model are estimated using the maximum likelihood estimation method, as shown in the formula:

[0031]

[0032] Where λ(x) is the intensity function of the Hawkes distribution, and L(T) is the maximum natural function of the distribution given that the distribution occurs at a given time.

[0033] In step S3, the overall user experience quality is quantitatively evaluated using the following formula:

[0034]

[0035] Where φ(i,j) is the propagation delay caused by user i requesting data j, and Sim(d m ,d n The similarity between the requested and returned data is denoted as . QoE(i,j,m,n) represents the user's experience quality for this request.

[0036] In step S4, the fairness coefficient is defined as follows:

[0037]

[0038] Where σ represents the standard deviation of the overall user experience quality calculated in S3, and H and L represent the upper and lower bounds of the overall user experience quality, respectively.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This invention utilizes a recommendation system to effectively mine user preferences and data features. By combining users' historical request data and employing techniques such as user-data rating matrix construction and non-negative matrix factorization, it can more accurately predict users' potential needs. At the same time, it introduces fairness indicators to quantitatively analyze the QoE differences among different users, ensuring that the caching strategy not only improves the overall service quality but also balances the experience gap between users in multi-user scenarios, thereby achieving higher user satisfaction.

[0041] 2. This invention combines recommendation systems with fairness analysis to optimize cache allocation by intelligently predicting user demand, while ensuring that all users receive a relatively fair quality of service when sharing resources. This approach is particularly suitable for service-intensive application scenarios, improving overall system performance while guaranteeing service balance among different users, thus providing more comprehensive protection for user experience in edge computing environments. Attached Figure Description

[0042] Figure 1 This is a flowchart of a fair-aware edge data caching method based on a recommendation system according to an embodiment of the present invention.

[0043] Figure 2 This describes the operation of the edge caching architecture of the recommendation system at a certain moment in an embodiment of the present invention. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0045] like Figure 1 As shown, a fair-aware edge data caching method based on a recommender system includes the following steps:

[0046] S1. Data Feature Extraction: Extracting data features from historical records for data value assessment and cache soft hit determination. The specific steps are as follows:

[0047] S11. User-Data Rating Matrix Construction: User request frequency is used as a rating, with higher request frequency resulting in a higher rating. A rating matrix M is constructed based on historical requests, where M... ij This indicates the number of times the user UI requests data (dj).

[0048] S12. Nonnegative Matrix Decomposition: The nonnegative matrix decomposition method is used to obtain the user feature matrix and the data feature matrix. Specifically, this algorithm attempts to find matrices P and Q such that the difference between the product of P and Q and M is minimized, i.e., minimizing:

[0049]

[0050] The specific update uses loss calculation and gradient descent to iterate continuously, that is...

[0051]

[0052] Ultimately, we can obtain the user feature matrix P and the data feature matrix Q for subsequent calculations.

[0053] S13, Matrix Iterative Update: An iterative update algorithm is used to optimize matrices P and Q.

[0054] S14. Data Similarity Calculation: The similarity between data is determined by calculating the cosine similarity between the feature vectors of the data, that is:

[0055]

[0056] S2. Predicting User Requests: A Hawkes process is used to predict future user requests, capturing temporal correlations to improve cache hit rate. The specific steps are as follows:

[0057] S21. Behavioral Correlation Analysis: Analyze the strong correlation between a user's historical behavior and future behavior. For example, if a user frequently requests a certain data recently, the user is more likely to continue requesting that data in the short term.

[0058] S22. Hawkes Process Modeling: User requests are modeled using Hawkes processes. A Hawkes process is a self-excited process in which the probability of an event occurring at a given time point is influenced by the occurrence of similar past events. In a Hawkes process, the probability of an event occurring at a given time is influenced by the occurrence of similar past events. The intensity function of time slot t can be expressed as:

[0059]

[0060] S23. Kernel Function Selection: This invention selects an exponentially decaying function as the kernel function, exhibiting monotonically decreasing characteristics to reflect the more significant impact of recent events on future request rates, i.e.:

[0061] φ(x)=αe -σ

[0062] Here, α and σ are the parameters of the function.

[0063] S24. Parameter Estimation and Calculation of Probability Density Function and Likelihood Function: The parameters in the Hawkes process model are estimated using the maximum likelihood estimation method. After constructing the maximum likelihood function, the likelihood function is iteratively optimized using the expectation-maximization (EM) algorithm, and the probability density function f(t) and the corresponding parameters of the Hawkes process are further derived.

[0064] S3. Maximizing User Experience Quality Allocation: This invention posits that Quality of Experience (QoE) depends on data retrieval latency and the similarity between retrieved and requested data. However, the relationship between these factors and QoE is often non-linear. For example, when the similarity falls below a certain threshold, the returned data will no longer match the user's expectations, leading to a significant decrease in QoE. Therefore, a sigmoid function is used to describe this relationship:

[0065]

[0066] Data similarity can be obtained in S1 when extracting data features and decomposing the feature matrix. Then, this invention designs a heuristic algorithm to optimize the caching strategy based on the predicted user requests obtained in S2: prioritizing cache pairs that maximize the overall experience quality until the cache space is full, aiming to find an approximate optimal solution that maximizes the overall user experience quality. The specific steps are as follows:

[0067] S31. Initialize data. Inputs: request set, cache block set, user set, edge server network G(S,E).

[0068] S32. First, initialize the gain wj of the cache block on the overall user experience quality of the system, and traverse each request in the request set to find the associated cache pair and update the gain of the cache block on the overall experience quality.

[0069] S33. Iterative selection of cache pairs: When the cache space is not full, continuously select the cache block with the maximum gain.

[0070] S34. Check cache space: If the cache space on the edge server is larger than the selected data size, cache the data on the server and update the server cache space.

[0071] S35. Update the system experience quality gain of the cache block: Iterate through the set of requests associated with the cache block and update the system experience quality gain value of each request in the current cache state.

[0072] S36. Repeat until the cache space is full: Repeat the above steps, continuously selecting the cache pair with the maximum gain, until the cache space is fully utilized.

[0073] S4. Cache Allocation Re-optimization: In step S4, when allocating cache to maximize user experience quality, the fairness of the caching strategy is first quantitatively evaluated based on the data feature information obtained in S1 and the user experience quality calculated in S3. This invention introduces a fairness index F to evaluate the fairness of the caching strategy. The fairness index F is defined as follows:

[0074]

[0075] Where σ represents the standard deviation of the overall user QoE calculated in S3, and H and L represent the upper and lower bounds of QoE, respectively. By minimizing the fluctuation of QoE (i.e., the standard deviation σ), this fairness index ensures that the caching strategy improves the overall user experience quality while minimizing the experience gap between different users. Subsequently, the caching strategy in S3 is optimized and adjusted based on the fairness index, thereby ensuring a high overall user experience quality while also considering the fairness of the experience among users. The specific steps are as follows:

[0076] S41. Select users with the lowest QoE: First, find the users with the lowest QoE, and try to gradually optimize the requests of users whose QoE is below the median to improve the fairness of the caching strategy.

[0077] S42. Find the cache pair with the minimum gain: Find the cache pair associated with the request to be optimized and select the cache pair with the minimum gain.

[0078] S43, Cache Replacement: Replaces the currently stored data in the cache with the requested data to improve QoE.

[0079] S44, Target Weighted Sum Check: Check whether the replacement has increased the overall value of the weighted target. If not, restore the data to its original state.

[0080] S45. Repeat the above process until all requests to be optimized are completed.

[0081] Recommender systems can not only uncover user and data characteristics but also significantly influence user request patterns. By integrating recommender systems into mobile edge data caching, the overall effectiveness of caching strategies is significantly improved. It's worth noting that the data requested by users is often not irreplaceable. When the requested data is not cached on nearby edge servers, recommending other highly relevant content can still meet the user's needs; this process is known as "soft hit."

[0082] Figure 2 This demonstrates the operation of an edge caching architecture for a recommendation system at a specific moment. When user u1 requests data d1, edge server s1 returns the cached data d1 to the user. For user u2 requesting data d2, the system can choose to retrieve data d1 from a nearby edge server (s1 or s3) or, depending on latency, from the cloud. If user u3 requests data d3, they have two options: 1) Direct cache hit: Retrieve data d3 from the cloud, which provides accurate content but has higher latency; 2) Soft cache hit: Accept data d1 cached on edge server s3 that is similar to data d3, thus reducing latency.

[0083] By employing this caching strategy supported by the recommendation system, this invention not only achieves significant results in improving user experience quality but also further reduces the overall system response latency. The concept of soft hits combines the advantages of recommendation systems and caching strategies, enabling the system to, to a certain extent, replace missed data in resource-constrained edge environments, thereby achieving optimal resource utilization.

[0084] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fairness-aware edge data caching method based on a recommender system, characterized in that, Includes the following steps: S1. Based on the historical request data of users obtained from the recommendation system, extract data features from the historical request data and calculate the similarity between the data features; S2. User requests are modeled using a Hawkes process; during the modeling process, an exponential decay function is selected as the kernel function, exhibiting monotonically decreasing characteristics, and the parameters in the Hawkes process model are estimated using the maximum likelihood estimation method; the Hawkes process model is used to predict future user requests. S3. Based on the data feature information obtained in S1, a quantitative evaluation of the overall user experience quality is performed. Then, based on the predicted user requests obtained in S2, a heuristic algorithm is used to optimize the caching strategy: prioritizing cache pairs that maximize the overall experience quality until the cache space is full, thereby finding an approximate optimal solution that maximizes the overall user experience quality. The specific steps are as follows: S31. Initialize data. Input: request set, cache block set, user set, edge server network G(S,E); S32. First, initialize the gain wj of the cache block on the overall user experience quality of the system, and traverse each request in the request set to find the associated cache pair and update the gain of the cache block on the overall experience quality. S33. Iterative selection of cache pairs: When the cache space is not full, continuously select the cache block with the maximum gain; S34. Check cache space: If the cache space on the edge server is larger than the selected data size, cache the data on the server and update the server cache space. S35. Update the system experience quality gain of the cache block: Iterate through the set of requests associated with the cache block and update the system experience quality gain value of each request in the current cache state. S36. Repeat until the cache space is full: Repeat the above steps, continuously selecting the cache pair with the maximum gain, until the cache space is fully utilized. The overall quality of the user experience is quantitatively evaluated using the following formula: Where φ(i,j) is the propagation delay caused by user i requesting data j, and Sim(d m ,d n ) represents the similarity between the requested and returned data; QoE(i,j,m,n) represents the user's experience quality for this request; S4. Based on the data feature information obtained in S1 and the user experience quality calculated in S3, a quantitative index of fairness coefficient is given, and the caching strategy in S3 is optimized and adjusted based on this quantitative index. The specific steps are as follows: S41. Select the user with the lowest QoE: First, find the user with the lowest QoE, and try to gradually optimize the requests of users whose QoE is lower than the median in order to improve the fairness of the caching strategy. S42. Find the cache pair with the minimum gain: Find the cache pair associated with the request to be optimized and select the cache pair with the minimum gain; S43, Cache Replacement: Replaces the currently stored data in the cache with the requested data to improve QoE; S44, Target Weighted Sum Check: Check whether the replacement has increased the overall value of the weighted target. If not, restore the data to its original state. S45. Repeat the above process until all requests to be optimized are completed. The fairness coefficient is defined as follows: Where σ represents the standard deviation of the overall user experience quality calculated in S3, and H and L represent the upper and lower bounds of the overall user experience quality, respectively.

2. The fair-aware edge data caching method based on a recommender system according to claim 1, characterized in that, The specific process of step S1 is as follows: The frequency of user requests is used as a score for the data, and a user-data score matrix M is constructed based on historical requests; The user-data rating matrix is ​​decomposed using a nonnegative matrix factorization algorithm to obtain the user feature matrix P and the data feature matrix Q. An iterative update algorithm is used to optimize the user feature matrix P and the data feature matrix Q; the similarity between the data is calculated.

3. The fair-aware edge data caching method based on a recommender system according to claim 2, characterized in that, The goal of the nonnegative matrix factorization algorithm is to find matrices P and Q such that the difference between the product of P and Q and M is minimized. The formula is: In the formula, Let M represent the sum of squares of the differences between all corresponding elements of matrix M and matrix PQ. ij This indicates the number of times user i requests data j.

4. The fair-aware edge data caching method based on a recommender system according to claim 2, characterized in that, An iterative update algorithm is used to optimize the user feature matrix P and the data feature matrix Q. Specifically, loss calculation and gradient descent are employed for iterative updates, as shown in the following formula: in, This represents the value in the i-th row and j-th column of matrix P in the N-th cycle. This represents the value in the i-th row and j-th column of matrix Q in the N-th cycle.

5. The fair-aware edge data caching method based on a recommender system according to claim 2, characterized in that, The similarity between data is calculated by using the cosine similarity between the feature matrices of the data, as shown in the following formula: In the formula, Sim(d m ,d n ) represents d m ,d n The similarity between them, cos(Q) m Q n Let |U| be the cosine similarity of the vectorized representations of the two data points. m ∩U n | represents the intersection of the user sets requesting these two data sets, and Θ is the confidence threshold to avoid high similarity due to too few records.

6. The fair-aware edge data caching method based on a recommender system according to claim 1, characterized in that, In step S2, during the Hawkes process, the probability of an event occurring at a given time is influenced by the occurrence of similar events in the past. The intensity function of time slot t is expressed as: Among them, H tn The set of timestamps represents historical events, β is the background intensity, and φ(t-τ) is the kernel function that captures the impact of past events on the arrival rate of similar events.

7. The fair-aware edge data caching method based on a recommender system according to claim 1, characterized in that, The parameters in the Hawkes process model are estimated using the maximum likelihood estimation method, as shown in the formula: Where λ(x) is the intensity function of the Hawkes distribution. Given a distribution that occurs at a given time, the largest natural function of its distribution.

Images