48th Transport Modellers Forum Meeting

18/04/2024

Tills, Tensors & Trajectories

​

A Location Planning Perspective on Mobility Data and  Machine Learning

​

AGENDA

​

• Location Planning

​

• Mobility Data

​

• Machine Learning

​

• Location Planning & Mobility Data & Machine Learning

2

ABOUT GEOLYTIX

WHAT WE DO

​

4

WHAT WE DO

​

5

WHO WE DO IT FOR

6

Mobility Data

MOBILITY DATA

who? 3fe33ed0-610a-11ed-9b6a-0242ac120002

when? 2022-11-10T15:14:34+00:00

where? 51.36855491432777, 0.3822254854173581

--------------------------------------------

source: 610a3ed0-610a-11ed-9b6a-11edac120002

horizontal deviation: 24m

...

• snapshots, not trajectories!
• use appropriate reference data to adjust long-term trends
• most characteristics of a mobility feed change over time
• but some do not:

8

Machine Learning

MODEL TYPES (SELECTION)

Increasing Complexity (Generally)

S Cubed ‘Spatial’

* Non-negative Least Squares

10

Risk mitigation in the context of small sample sizes:

​

• Oversampling: When dealing with imbalanced data (more data for some store formats than others), we use oversampling to create a more balanced training set.

​

• Stratified Splitting: We avoid random splits for training and testing. Instead, we use stratified sampling to ensure a good spread of different store formats and location types in both sets.

​

• Hyperparameter Tuning: We tune hyperparameters using techniques like random search with cross-validation to optimize model performance.

ML CHALLENGES AND SOLUTIONS

SAMPLE SIZE

RISK

Healthcare

Financial Services

Location Planning

CRM

A/B Testing

Scientific Research

11

​

​

INTERACTION MODELS VS “FLAT” MODELS

• interaction models: stores (own and competition) interact on a demand level

​

• flat models: store x feature matrix

​

• manual scorecards

​

• (constrained) regression models

​

• gradient boosting algorithms, e.g. XGBoost, LightGBM

​

​

• decision usually driven by whether external/transient factors deemed more salient than interaction between stores/competitors

​

vs

12

1. Gravity Models
2. Market Share ML Models

​

• we model the flows of consumer demand from the smallest level of administrative geography to Retail Places or individual stores

​

• Summing up these flows gives a robust and objective means of forecasting potential sales

​

• The key underlying principle is consumers tend to behave rationally and select destinations that are more convenient and more attractive (greater choice)

​

SPATIAL INTERACTION MODELS

13

Where Machine Learning & Mobility Data (can) meet

• Making meaningful input variables, rather than letting an ML model figuring out biases and coverage issues, usually works out better for us (GI/GO)

​

• Consistency between what clients see and what goes into their models

​

​

INPUT VARIABLES

Footfall

Traffic

15

• Calculating catchments helps us understand the role of a store/town and its attractiveness, how it overlaps with nearby stores/towns

​

• Using customer data and/or modelled catchment data we map the flow of customer spend allowing us to generate ‘natural’ primary and secondary catchments

​

• Catchment creation and analysis plays a crucial role in optimising a store network

​

• The Retail Place type, catchment extents and coverage (household, population, demand, target customers) can all be used to help determine the importance of having a store in a given area

CATCHMENTS CREATION

16

• Distance decay reflects how far people are willing to travel to a Retail Place, recognising that people in rural areas are often more willing to travel further than those people living in a city centre, whose provisions are typically met within a small distance from home

​

• Our distance decay values are defined at OA level using the urbanity score of the OA

​

DISTANCE DECAY MODELLING

​

17

• Example: catchment overlap measure

​

• As a standalone model, or

​

• inbuilt in spatial components of feature matrix

​

• leave transient model components untouched

RESIDENTIAL IMPACTS

NEW STORE

Nearest Catchments

Nearest Catchments

18

• Interaction Surfaces Product

​

• Probability of a person showing up in other locations on the same day

​

• Measure of shared transient population potential

TRANSIENT IMPACTS

19

NETWORK BLUEPRINTS

20

The Future

site vs rest of the world (flat model)

​

site vs anything else (interaction model)

​

anything vs anything

WHAT’S NEXT

Top-down

​

(e.g. Graph Neural Nets)

Bottom-up

​

(e.g. Causally informed synthetic populations)

22

48th Transport Modellers Forum Meeting

18/04/2024

THANK YOU