Improved Reverse Geocoding - Context Detection Project
At Sony Techsoft Centre, we are conducting R&D in the area of context detection and data analysis. Our goal is to enable new user experiences through analysis of data gathered via the user’s devices, and by enriching this user data with other information, for example from third-party services.
Currently we focus on data available from Android devices such as mobile phones and wearables, specifically
- Location data (GPS, Google location)
- Activity data (walking, cycling, in vehicle, …)
- Network data (WiFi, Bluetooth connection & scan, cell tower, …)
- Labelled Ground truth about visited locations
We have gathered a sizable annotated data set (60+ users, for several months) that is ready for analysis, and continue growing and diversifying it.
- The goal of this project is to detect the places where people spend time and label those places as accurately as possible.
- We distinguish in 2 categories of places:
- public places such as restaurants, airports, shops, concert halls, sports clubs etc.
- private places such as home, friends homes, etc.
- The output of this project will be used to create enriched user profiles and help in detecting their interests and habits.
- A simple approach will use simple reverse geocoding using online services. But we need to overcome inaccurate GPS fixes when indoors could be 100m off. So the intention is to use habit detection and machine learning to improve the accuracy
- In a competition in 2012 based on the Nokia dataset, the best of class algorithms achieved between 65% and 75% prediction accuracy
- Establish a baseline prediction rate using regular reverse geocoding systems (google, open streetmap,..).
- Come up with an improvement strategy and implement it.
- The Sony Lynx dataset has ground truth + GPS, wifi, activity data for 60+ users (multiple weeks),
- Puget sound datasets (http://psrc.github.io/2014/hh-survey-pr1/ , https://www.icpsr.umich.edu/icpsrweb/ICPSR/series/263/studies/35350?paging.startRow=1)