Mobile Edge Cache Prediction with Federated Learning and Attention Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing mobile edge caching methods in MEC environments face issues with high terminal equipment consumption due to all users participating in training, uneven local model quality leading to slow global model convergence, and inefficient cache hit rates due to user mobility and content popularity changes.
Innovation Solution
A federated learning approach is used to select users for training based on clustering and threshold selection, combined with an attention mechanism for model aggregation, predicting content popularity and caching it proactively to improve cache hit rates and reduce local consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all users participate in federated learning training within a base station, then the global model can be trained with diverse data, but the terminal equipment consumption increases unnecessarily
Solution Approach 1:
The patent segments users into different clusters based on their request patterns and selects representative users from each cluster to participate in training. This segmentation allows the system to maintain diverse data representation while reducing the number of active participants, thereby lowering terminal energy consumption without compromising global model quality.
Solution Approach 2:
Instead of requiring all users to participate in training, the patent applies partial action by selecting only a subset of users (specifically, representative users from each cluster) to participate. This partial participation approach reduces overall energy consumption while still achieving sufficient model training diversity through strategic user selection.
2Reliability
If simple average aggregation is used to aggregate local model parameters, then the global model can be obtained, but the convergence speed slows down due to uneven local model quality
Solution Approach 1:
The patent applies local quality by evaluating and weighting local models based on their specific characteristics and contributions. Instead of treating all local models equally, the system adjusts aggregation weights according to the quality and relevance of each local model, thereby accelerating convergence while maintaining model accuracy.
Solution Approach 2:
The patent changes the aggregation parameter from simple average to weighted average, where weights are dynamically adjusted based on local model quality metrics. This parameter change enables the system to prioritize high-quality local models during aggregation, significantly improving convergence speed without sacrificing model reliability.
3Measurement precision
If centralized training of all user data is performed, then content popularity can be predicted accurately, but network load increases and user privacy is threatened
Solution Approach 1:
The patent segments the training process into distributed local training and centralized aggregation phases. Each user trains locally on their own data, preventing large-scale data transmission. Only model parameters are exchanged between users and the server, significantly reducing network load while maintaining prediction accuracy through collaborative learning.
Solution Approach 2:
The patent introduces model parameters as an intermediary representation of user data. Instead of transmitting raw user data to the server, the system uses locally trained model parameters as intermediaries that capture essential patterns. This intermediary approach enables accurate popularity prediction while protecting user privacy and reducing network bandwidth requirements.
Data Source
AI summary
The invention relates to an optimization method for mobile edge cache based on federated learning. The situation that the user mobility and the content popularity change continuously in the range of a single base station is considered, and the cache hit rate is increased by predicting the content popularity and placing the request content in an edge cache in advance. The method includes obtaining a user moment trajectory table to simulate a moving path of a user by using an RWP random path point model, selecting the user participating in FL local training in a clustering, and threshold value combination mode in consideration of local training consumption, performing global model aggregation by using an attention mechanism to control model weight, and performing global prediction according to an obtained global prediction model. The predicted request content is cached to the server in advance.