Telecom Cell Clustering for Data-Scarce Machine Learning Collaboration

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Mobile cellular networks face challenges in data scarcity for machine learning-driven Self-Organizing Network (SON) functions, leading to inefficient model training and performance degradation due to sparse labeled data, which existing solutions like interpolation techniques and Generative Adversarial Networks (GANs) are inadequate for non-overlapping latent features.

Innovation Solution

A method is provided to identify and facilitate collaboration between SON models in mobile networks through transfer learning by extracting and clustering cells based on similarities in clutter distribution, training dataset meta-features, and model hyper-parameters, using a collaboration server to optimize knowledge sharing among cells.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If machine learning models are trained using local data only, then model training can be performed independently at each cell, but data scarcity and sparse labeled data lead to inefficient model training and performance degradation

Engineering Contradiction:
Improvemodel training efficiencyVSAvoidamount of training data
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent combines training data from multiple cells through data sharing mechanisms. Cells with similar characteristics (grouped by clustering algorithms) share their labeled data, effectively merging limited local datasets into larger training sets. This resolves the data scarcity problem by pooling resources across the network while maintaining independence through selective sharing based on similarity metrics.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent introduces a data sharing platform or intermediary system that facilitates controlled data exchange between cells. This intermediary manages the clustering, similarity assessment, and data transfer processes, enabling cells to access external training data without direct peer-to-peer communication. The intermediary resolves the contradiction by mediating between local data limitations and the need for sufficient training data.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If cells share training data and models across the network, then model performance improves, but signaling costs and network complexity increase

Engineering Contradiction:
Improvemodel performanceVSAvoidsignaling costs
Core Design Contradiction:
ReliabilityVSLoss of energy

Solution Approach 1:

The patent applies local quality by enabling data sharing only between cells with similar characteristics rather than universal sharing. Clustering algorithms identify groups of cells with comparable traffic patterns, geographic features, and performance metrics. Only within these similar groups does data sharing occur, reducing unnecessary signaling while maintaining performance benefits. This selective approach resolves the contradiction between performance improvement and signaling cost reduction.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent segments the network into multiple clusters based on cell characteristics, with each cluster independently managing its own data sharing. This segmentation prevents full-network signaling overhead by limiting data exchange to relevant subsets. Each cluster operates semi-independently, reducing the overall signaling burden while still providing performance benefits through localized knowledge sharing.

Inventive Principle:
Principle #1Segmentation

3Quantity of substance

If interpolation techniques and GANs are used to generate training data, then data availability increases, but these methods are inadequate for non-overlapping latent features

Engineering Contradiction:
Improveavailability of training dataVSAvoidaccuracy of generated data
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent performs preliminary clustering and similarity assessment before data sharing or generation. By pre-identifying cells with overlapping latent features through feature analysis and clustering, the system ensures that data sharing or GAN generation only occurs when feature compatibility exists. This preliminary action prevents the generation of inaccurate data for non-overlapping features while still providing sufficient training data where appropriate.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses real data copying from similar cells rather than synthetic generation for cells with overlapping latent features. When clustering identifies cells with comparable characteristics, the system copies actual labeled data from one cell to another, preserving the authenticity and accuracy of the training data. This copying approach avoids the accuracy limitations of GANs while still increasing data availability for cells with insufficient local data.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4226589B1Method for identifying potential machine learning model candidates to collaborate in telecom networks
Publication Date: 2025.09.03 TELEFONAKTIEBOLAGET LM ERICSSON (PUBL)
  • EP4226589B1 patent drawingFigure 1
  • EP4226589B1 patent drawingFigure 2
  • EP4226589B1 patent drawingFigure 3

AI summary

A method of a collaboration server for identifying cells of a mobile network for machine learning collaboration is provided. The mobile network includes a plurality of cells. The method includes managing collection of features for the plurality of cells to generate at least one feature vector for each of the plurality of cells, determining a cluster of cells within the plurality of cells based on similarity in feature vectors between at least two cells in the plurality of cells, sending cluster information to each cell of the cluster, receiving cluster pre-check information from each cell of the cluster, and determining a first cell and a second cell in the cluster to collaborate for machine learning based on the received pre-check information.