Cellular Geolocation Training Data Grids for ML Bias Reduction

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning (ML) models for geolocation prediction in cellular communication systems face issues with biased and disproportionate geo-spatial distribution of training data, leading to excessive data size and resource burden.

Innovation Solution

A method is introduced to define grids with grid boxes covering geographic areas, limiting the number of truth call records per box to control the training data distribution, replacing older records when the limit is reached, and using configurable parameters for larger grids in extensive areas to manage bias and reduce data size.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If truth data from all UEs is collected for training ML models, then the quantity of training data increases, but the geo-spatial distribution becomes disproportionate and biased

Engineering Contradiction:

Improvequantity of training dataVSAvoidgeo-spatial distribution uniformity

Core Design Contradiction:

Quantity of substanceVSManufacturing precision

Solution Approach 1:

The geographic area is divided into multiple grid boxes, and truth data is sampled independently from each grid box. This segmentation allows controlled representation from different geographic regions, ensuring proportional geo-spatial distribution while maintaining sufficient data quantity for training.

Inventive Principle:

Principle #1Segmentation

2Quantity of substance

If truth data from the same UE or cluster is collected extensively, then the quantity of training data increases, but bias is introduced in the ML model

Engineering Contradiction:

Improvequantity of training dataVSAvoidmodel accuracy

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

Different grid boxes have independent sampling limits, allowing the system to control the contribution from each local geographic area. This prevents any single UE or cluster from dominating the training data, thereby reducing bias and improving model reliability while maintaining overall data quantity.

Inventive Principle:

Principle #3Local quality

3Manufacturing precision

If maximum number of entries per grid box is limited, then bias is reduced in training data, but the quantity of training data decreases

Engineering Contradiction:

Improvegeo-spatial distribution uniformityVSAvoidquantity of training data

Core Design Contradiction:

Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent introduces a spatial dimension (grid boxes) to organize training data, transforming the problem from a single pooled dataset to a multi-dimensional structured dataset. This allows independent control of data quantity per geographic region while maintaining sufficient overall data volume for effective model training.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Data Source

PatentUS12395590B2Reduction and geo-spatial distribution of training data for geolocation prediction using machine learning

Publication Date: 2025.08.19 NETSCOUT SYSTEMS TEXAS LLC

AI summary

A method is provided of limiting machine learning (ML) model training data, including receiving a maximum number of entries per grid box N. Grids are defined, each grid having grid boxes and covering a geographic area defined by a different cell list of a communication network, each grid box covering a different portion of the geographic area. Truth call records having truth data reporting geolocation (GL) data are selected from multiple received call records. A grid box is determined for each truth call record that covers a geographic area including the GL indicated by the truth call record's GL data. Truth call records are selectively assigned to their determined grid box in a fashion to not exceed the maximum number of entries per grid box N, and the truth data and signal detail data of only the assigned truth call records are output as the ML model training data.