Generating synthetic mobility data for a realistic population with RNNs to improve utility and privacy

Overview

lbs-data

Motivation

Location data is collected from the public by private firms via mobile devices. Can this data also be used to serve the public good while preserving privacy? Can we realize this goal by generating synthetic data for use instead of the real data? The synthetic data would need to balance utility and privacy.

Overview

What:

This project uses location based services (LBS) data provided by a location intelligence company in order to train a RNN model to generate synthetic location data. The goal is for the synthetic data to maintain the properties of the real data, at the individual and aggregate levels, in order to retain its utility. At the same time, the synthetic data should sufficiently differ from the real data at the individual level, in order to preserve user privacy.

Furthermore, the system uses home and work areas as labels and inputs in order to generate location data for synthetic users with the given home and work areas.
This addresses the issue of limited sample sizes. Population data, such as census data, can be used to create the input necessary to output a synthetic location dataset that represents the true population in size and distribution.

Data

/data/

ACS data

data/ACS/ma_acs_5_year_census_tract_2018/

Population data is sourced from the 2018 American Community Survey 5-year estimates.

LBS data

/data/mount/

Privately stored on a remote server.

Geography and time period

  • Geography: The region of study is limited to 3 counties surrounding Boston, MA.
  • Time period: The training and output data is for the first 5-day workweek of May 2018.

Data representation

The LBS data are provided as rows.

device ID, latitude, longitude, timestamp, dwelltime

The data are transformed into "stay trajectories", which are sequences where each index of a sequence represents a 1-hour time interval. Each stay trajectory represents the data for one user (device ID). The value at that index represents the location/area (census tract) where the user spent the most time during that 1-hour interval.

e.g.

[A,B,D,C,A,A,A,NULL,B...]

Where each letter represents a location. There are null values when no location data is reported in the time interval.

home and work locations are inferred for each user stay trajectory. stay trajectories are prefixed with the home and work locations. This home, work prefixes then serve as labels.

[home,work,A,B,D,C,A,A,A,NULL,B...]

Where home,work values are also elements (frequently) occuring in their associated stay trajectory (e.g. home=A).

These sequences are used to train the model and are also output by the model.

RNN

The RNN model developed in this work is meant to be simple and replicable. It was implemented via the open source textgenrnn library. https://github.com/minimaxir/textgenrnn.

Many models (>70) are trained with a variety of hyper parameter values. The models are each trained on the same training data and then use the same input (home, work labels) to generate output synthetic data. The output is evalued via a variety of utility and privacy metrics in order to determine the best model/parameters.

Pipeline

Preprocessing

Define geography / shapefiles

./shapefile_shaper.ipynb

Our study uses 3 counties surrounding Boston, MA: Middlesex, Norfolk, Suffolk counties.

shapefile_shaper prunes MA shapefiles for this geography.

Output files are in ./shapefiles/ma/

Census tracts are used as "areas"/locations in stay trajectories.

Data filtering

./preprocess_filtering.ipynb

The LBS data is sparse. Some users report just a few datapoints, while other users report many. In order to confidently infer home and work locations, and learn patterns, we only include data from devices with sufficient reporting.

./preprocess_filtering.ipynb filters the data accordingly. It pokes the data to try to determine what the right level of filtering is. It outputs saved files with filtered data. Namely, it saves a datafile with LBS data from devices that reported at least 3 days and 3 nights of data during the 1 workweek of the study period. This is the pruned dataset used in the following work.

Attach areas

/attach_areas.ipynb

Census areas are attached to LBS data rows.

Home, work inference

./infer_home_work.ipynb

Defines functions to infer home and work locations (census tracts ) for each device user, based on their LBS data. The home location is where the user spends most time in nighttime hours. The "work" location is where the user spends the most time in workday hours. These locations can be the same.

This file helps determine good hours to use for nighttime hours. Once the functions are defined, they are used to evaluate the data representativeness by comparing the inferred population statistics to ACS 2018 census data.

Saves a mapping of LBS user IDS to the inferred home,work locations.

Stay trajectories setup

./trajectory_synthesis/trajectory_synthesis_notebook.ipynb

Transforms preprocessed LBS data into prefixed stay trajectories.

And outputs files for model training, data generation, and comparison.

Note: for the purposes of model training and data generation, the area tokens within stay trajectories can be arbitrary. What is important for the model’s success is the relationship between them. In order to save the stay trajectories in this repository yet keep real data private, we do the following. We map real census areas to integers, and map areas in stay trajectories to the integers representing the areas. We use the transformed stay trajectories for model training and data generation. The mapping between real census areas and their integer representations is kept private. We can then map the integers in stay trajectories back to the real areas they represent when needed (such as when evaluating trip distance metrics).

Output files:

./data/relabeled_trajectories_1_workweek.txt: D: Full training set of 22704 trajectories

./data/relabeled_trajectories_1_workweek_prefixes_to_counts.json: Maps D home,work label prefixes to counts

./data/relabeled_trajectories_1_workweek_sample_2000.txt: S: Random sample of 2000 trajectories from D.

./data/relabeled_trajectories_1_workweek_prefixes_to_counts_sample_2000.json: Maps S home,work label prefixes to counts

  • This is used as the input for data generation so that the output sythetic sample, S', has a home,work label pair distribution that matches S.

Model training and data generation

./trajectory_synthesis/textgenrnn_generator/

Models with a variety of hyperparameter combinations were trained and then used to generate a synthetic sample.

The files model_trainer.py and generator.py are the templates for the scripts used to train and generate.

The model (hyper)parameter combinations were tracked in a spreadsheet. ./trajectory_synthesis/textgenrnn_generator/textgenrnn_model_parameters_.csv

Evaluation

./trajectory_synthesis/evaluation/evaluate_rnn.ipynb

A variety of utility and privacy evaluation tools and metrics were developed. Models were evaluated by their synthetic data outputs (S'). This was done in ./trajectory_synthesis/evaluation/evaluate_rnn.ipynb. The best model (i.e. best parameters) was determined by these evaluations. The results for this model are captured in trajectory_synthesis/evaluation/final_eval_plots.ipynb.

Owner
Alex
Systems Architect, product oriented Engineer, Hacker for the social good, Math Nerd that loves solving hard problems and working with great people.
Alex
✂️ EyeLipCropper is a Python tool to crop eyes and mouth ROIs of the given video.

EyeLipCropper EyeLipCropper is a Python tool to crop eyes and mouth ROIs of the given video. The whole process consists of three parts: frame extracti

Zi-Han Liu 9 Oct 25, 2022
"SOLQ: Segmenting Objects by Learning Queries", SOLQ is an end-to-end instance segmentation framework with Transformer.

SOLQ: Segmenting Objects by Learning Queries This repository is an official implementation of the paper SOLQ: Segmenting Objects by Learning Queries.

MEGVII Research 179 Jan 02, 2023
Jittor implementation of PCT:Point Cloud Transformer

PCT: Point Cloud Transformer This is a Jittor implementation of PCT: Point Cloud Transformer.

MenghaoGuo 547 Jan 03, 2023
Trajectory Prediction with Graph-based Dual-scale Context Fusion

DSP: Trajectory Prediction with Graph-based Dual-scale Context Fusion Introduction This is the project page of the paper Lu Zhang, Peiliang Li, Jing C

HKUST Aerial Robotics Group 103 Jan 04, 2023
Applicator Kit for Modo allow you to apply Apple ARKit Face Tracking data from your iPhone or iPad to your characters in Modo.

Applicator Kit for Modo Applicator Kit for Modo allow you to apply Apple ARKit Face Tracking data from your iPhone or iPad with a TrueDepth camera to

Andrew Buttigieg 3 Aug 24, 2021
DumpSMBShare - A script to dump files and folders remotely from a Windows SMB share

DumpSMBShare A script to dump files and folders remotely from a Windows SMB shar

Podalirius 178 Jan 06, 2023
Ranger - a synergistic optimizer using RAdam (Rectified Adam), Gradient Centralization and LookAhead in one codebase

Ranger-Deep-Learning-Optimizer Ranger - a synergistic optimizer combining RAdam (Rectified Adam) and LookAhead, and now GC (gradient centralization) i

Less Wright 1.1k Dec 21, 2022
Tgbox-bench - Simple TGBOX upload speed benchmark

TGBOX Benchmark This script will benchmark upload speed to TGBOX storage. Build

Non 1 Jan 09, 2022
Official Implementation of VAT

Semantic correspondence Few-shot segmentation Cost Aggregation Is All You Need for Few-Shot Segmentation For more information, check out project [Proj

Hamacojr 114 Dec 27, 2022
A torch.Tensor-like DataFrame library supporting multiple execution runtimes and Arrow as a common memory format

TorchArrow (Warning: Unstable Prototype) This is a prototype library currently under heavy development. It does not currently have stable releases, an

Facebook Research 536 Jan 06, 2023
The codes for the work "Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation"

Swin-Unet The codes for the work "Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation"(https://arxiv.org/abs/2105.05537). A validatio

869 Jan 07, 2023
AdelaiDepth is an open source toolbox for monocular depth prediction.

AdelaiDepth is an open source toolbox for monocular depth prediction.

Adelaide Intelligent Machines (AIM) Group 743 Jan 01, 2023
Unity Propagation in Bayesian Networks Handling Inconsistency via Unity Smoothing

This repository contains the scripts needed to generate the results from the paper Unity Propagation in Bayesian Networks Handling Inconsistency via U

0 Jan 19, 2022
A high-level Python library for Quantum Natural Language Processing

lambeq About lambeq is a toolkit for quantum natural language processing (QNLP). Documentation: https://cqcl.github.io/lambeq/ Getting started Prerequ

Cambridge Quantum 315 Jan 01, 2023
Matlab Python Heuristic Battery Opt - SMOP conversion and manual conversion

SMOP is Small Matlab and Octave to Python compiler. SMOP translates matlab to py

Tom Xu 1 Jan 12, 2022
The Official Implementation of the ICCV-2021 Paper: Semantically Coherent Out-of-Distribution Detection.

SCOOD-UDG (ICCV 2021) This repository is the official implementation of the paper: Semantically Coherent Out-of-Distribution Detection Jingkang Yang,

Jake YANG 62 Nov 21, 2022
[NeurIPS-2021] Slow Learning and Fast Inference: Efficient Graph Similarity Computation via Knowledge Distillation

Efficient Graph Similarity Computation - (EGSC) This repo contains the source code and dataset for our paper: Slow Learning and Fast Inference: Effici

24 Dec 31, 2022
Learning to Disambiguate Strongly Interacting Hands via Probabilistic Per-Pixel Part Segmentation [3DV 2021 Oral]

Learning to Disambiguate Strongly Interacting Hands via Probabilistic Per-Pixel Part Segmentation [3DV 2021 Oral] Learning to Disambiguate Strongly In

Zicong Fan 40 Dec 22, 2022
OpenABC-D: A Large-Scale Dataset For Machine Learning Guided Integrated Circuit Synthesis

OpenABC-D: A Large-Scale Dataset For Machine Learning Guided Integrated Circuit Synthesis Overview OpenABC-D is a large-scale labeled dataset generate

NYU Machine-Learning guided Design Automation (MLDA) 31 Nov 22, 2022
Kaggle | 9th place (part of) solution for the Bristol-Myers Squibb – Molecular Translation challenge

Part of the 9th place solution for the Bristol-Myers Squibb – Molecular Translation challenge translating images containing chemical structures into I

Erdene-Ochir Tuguldur 22 Nov 30, 2022