Explainability-Guided Model Aggregation for Faster Federated Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional federated learning approaches for predictive models in telecommunications networks suffer from sub-optimal performance and slower convergence due to the dominance of local models with more samples, leading to increased computational requirements and energy inefficiency, without effectively leveraging the attributes of well-performing local models to update the global model.
Innovation Solution
A method that utilizes explainability values, such as SHAP values, from well-performing local models to update the global model, allowing for faster convergence and improved performance by excluding models with degraded performance, thereby reducing computational iterations and enhancing energy efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional Federated Averaging is used to aggregate local models, then the global model can be updated, but the convergence is slow and performance is sub-optimal
Solution Approach 1:
The patent changes the aggregation parameter from simple weighted average (FedAvg) to a performance-based selection mechanism. Instead of aggregating all local models equally, the system identifies well-performing local models (those exceeding a performance threshold) and uses only their parameters for global model updates, fundamentally changing how aggregation weights are assigned.
Solution Approach 2:
The patent extracts only the useful components from local models - specifically, it extracts parameters from local models that meet the performance threshold criterion. Models with degraded performance are excluded from the aggregation process, extracting only high-quality contributions toward the global model.
2Quantity of substance
If all local models are used for aggregation regardless of performance, then more data is utilized, but computational requirements increase and energy efficiency decreases
Solution Approach 1:
The patent extracts only high-quality local models that meet the performance threshold, excluding degraded models from the aggregation process. This selective extraction reduces the total computational workload while maintaining effective data utilization from the most informative sources.
Solution Approach 2:
The patent applies different treatment to different local models based on their quality. Well-performing models (above threshold) are included in aggregation, while degraded models (below threshold) are excluded. This local differentiation optimizes computational resources by focusing only on high-quality contributions.
3Reliability
If drift detection and resolution mechanisms are implemented at local nodes, then model accuracy is maintained, but computational cost increases
Solution Approach 1:
The patent implements a feedback mechanism where local models report their performance metrics to the global node. Based on this feedback, the global node determines which local models meet the performance threshold and should contribute to aggregation. This feedback loop maintains accuracy while simplifying the overall computational burden compared to local drift detection mechanisms.
Data Source
AI summary
A computer-implemented method, performed by a first node (111), for handling predictive models. The first node (111) updates (207), using machine learning, a first predictive model of an indicator of performance of the communications system (100). The updating (207) is based on respective explainability values respectively obtained from a first subset of a plurality of second nodes (112). The respective explainability values correspond to a first subset of respective second predictive models of the indicator of performance of the communications system (100), respectively determined by the first subset of the plurality of second nodes (112). The models in the first subset of respective second predictive models have a respective performance value above a threshold. The first node (111) then provides (208) an indication of the updated first predictive model to a third node (113) comprised in the plurality of second nodes (112) and excluded from the first subset, or to another node (114).