[Pycon] [new paper] "Breno Freitas" - Adaptive Online Geospatial Machine Learning

[info a pycon.it]

Title: Adaptive Online Geospatial Machine Learning
Duration: 45 (includes Q&A)
Q&A Session: 15
Language: en
Type: Talk

Abstract: Uncovering points of interests (POI) is a common problem that companies try to solve. In Brazil, some companies are trying to track and trace bus drivers’ behaviours in order to avoid fraudsters that try not to pay the full price of a ride. One of the major issues with that is identifying POI for bus stops, stations and garages on the fly, given that inserting all of them manually is a cumbersome job that might required hours of work.

There are many different algorithms that one could use to adaptively find clusters of points given the interactions with buses along the time; However, GPS offers often unreliable and imprecise points that change over time and might not have the proper density for a clustering algorithm to work. On top of that, clusters might overlap over time and we want to always have them up to date.

In this talk we introduce a variation of a clustering algorithm, based on DBSCAN that adapts over time using a Graph Theory interpretation over geospatial overlaps, trained with online data receiving thousands of requests per minute over its API.

Tags: [u'machine-learning', u'geospatial']