Research Interests

My name is Mathieu Guillame-Bert, and I am working in Google Zurich, Switzerland.

I did my Postdoctoral at CMU (Carnegie Mellon University) in the Auton-Lab (part of Robotic Institute) under the supervision of Artur Dubrawski. I studied the use of rule extraction and decision forest-like techniques for symbolic and numerical time series. Typical applications include Automated and/or assisted Medical Diagnostics, Medical Forecasting, Human Activity Recognition and Forecasting, Banking Activity Patterns, Fault Detection and Prediction on complex systems, automatic Foreign exchange market trading, and a bit of Robotics.

I defended my PhD in France in 2012 at the INRIA Research Lab in the PRIMA team under the supervision of James L. Crowley. I studied the automated extraction of patterns from large temporal datasets, as well as the use of those models for automated prediction, user interpretation and automated reasoning.

I graduate from the Master of Advance Computing of the Imperial College of London, and from the "French Grande Ecole" ENSIMAG (Ecole Nationale Supérieure d'Informatique et de Mathématiques Appliquées) in 2009.

Publications

2018

Article

▪ Exact Distributed Training: Random Forest with Billions of Examples

M. Guillame-Bert and O. Teytaud.

Arxiv

[paper ]

2017

Journals/Conference Proceedings

▪ Classification of Time Sequences using Graphs of Temporal Constraints

M. Guillame-Bert and A. Dubrawski.

Journal of Machine Learning Research 18 (2017)

[paper ][algorithm sources ][pig dataset ]

2016

Journals/Conference Proceedings

Article

▪ Batched Lazy Decision Trees

ArXiv.org repository.

[arxiv ]

▪ Honey: A dataflow programming language for the processing, featurization and analysis of multivariate, asynchronous and non-uniformly sampled scalar symbolic time sequences

ArXiv.org repository.

[arxiv ]

Tutorial

▪ Honey programming language tutorial

Introduction to the Honey programming language which is designed to efficiently process time structured datasets.

[html ]

▪ Event Viewer video tutorial

Introduction to the Event Viewer software which is designed to visualise large time structured datasets.

[html ]

2015

Presentation

2014

Workshop

Presentation

2013

Presentation

▪ Learning Temporal Rules to Forecast Instability in Intensive Care Patients

Mathieu Guillame-Bert, Artur Dubrawski, Karen Chen, Andre Holder, Gilles Clermont, Marilyn Hravnak, and Michael Pinsky

ESICM 2013 (European Society of Intensive Care Medicine)

[read ]

Tutorial

▪ The TITARL algorithm

An interactive tutorial to understand the TITARL algorithm (Data-Mining algorithm on symbolic time sequences).

[read ]

2012

Journals/Conference Proceedings

▪ Learning Temporal Association Rules on Symbolic Time Sequences

Mathieu Guillame-Bert and James L. Crowley

In Proceedings of the 2012 4th Asian Conference on Machine Learning, Singapore, 2012

[read ]

▪ Planning with Inaccurate Temporal Rules

Mathieu Guillame-Bert and James L. Crowley

In Proceedings of the 2012 IEEE 24rd International Conference on Tools with Artificial Intelligence, Athens, Greece, 2012

[read ]

Presentation

▪ Learning Temporal Association Rules on Symbolic Time Sequences

Extract of the thesis defense

[read ]

Report

▪ PhD's Thesis, Learning Temporal Association Rules on Symbolic Time Sequences

Under the supervision of Pr. James L. CROWLEY

PRIMA Team – INRIA Lab. – Grenoble - France, 2012
Committee: Pr. Malik Gha llab, Pr. Paul Lukowicz, Dr. Artur Dubrawski, Pr. Augustin Lux

[read ]

2011

Journals/Conference Proceedings

▪ New Approach on Temporal Data Mining for Symbolic Time Sequences: Temporal Tree Associate Rules

Mathieu Guillame-Bert and James L. Crowley

In Proceedings of the 2011 IEEE 23rd International Conference on Tools with Artificial Intelligence

[read ]

▪ Predicting Home Service Demands from Appliance Usage Data

Kaustav Basu, Mathieu Guillame-Bert, Hussein Joumaa, Stephane Ploix and James Crowley

In Proceedings of the 3rd International Conference on Information and Communication Technologies and Applications ICTA 2011

[read ]

2010

Journals/Conference Proceedings

▪ First-order Logic Learningin Artificial Neural Networks

Mathieu Guillame-Bert, Krysia Broda and Artur d'Avila Garcez

In Proceedings of 23rd International Joint Conference on Neural Networks IJCNN 2010

[read ]

Prior to 2010

Report

▪ Master's Thesis, Connectionist Artificial Neural Network

With the supervision of Krysia Broda

Imperial College of London, 2009
Distinguished MSc project [page]

[read ]

▪ Bs.C work, I-Terms unification

With the supervision of Nicolas Peltier

ENSIMAG, 2008
Download implementation (C++) [.zip]

[read ]

Curriculum vitae

Softwares & Datasets

Click on this link bellow to see the softwares and the data-sets that I created or processed.

Softwares

• TITAR Learner
  TITAR (Temporal Interval Tree Association Rule Learner) is Temporal Data Mining algorithm able to extract temporal patterns from symbolic time series and time sequences. These patterns can then be used to user interpretation or for automatic forecasting.
• Event Viewer
  A powerful Time Sequences visualier. Event Viewer has been developed during the study of several real world problems. For that reason Event Viewer has a lot of unique features which allow a powerful understanding of data.
• Event Script
  Event script is a small but high level flow-oriented declarative programming language designed to facilitate the pre/post processing and the analysis of symbolic and numerical time series and sequences datasets.
• Machine Learning Lab.
  A small exploratory tool designed to test and experiment easily with several Machine Learning algorithms on several real work and sythetic datasets.

Public datasets

• The computer generated Database
  A syntetic data-set create to test symbolic time sequence data-mining algorithms.
• The Home Data Set
  The "Home activities dataset" of Tim van Kasteren et al. converted to be imported in TITARL and Event Viewer.
• Forex Dataset: Automated Trading Experiment
  A three years record of the USD/EUR Forrex market converted to be imported in TITARL and Event Viewer.

Gallery

This sections shows some of the nice pictures that I have generated through my research work. You can click on the picture to get the full size versions.

Learning 3d ellipse with k-nearest neighbor (KNN)

TITARL web interface

Example of our ploting library

Virtual robot in action (close view)

Learning 3d ellipse with Normal Bayes

Learning 3d ellipse with decision trees

Example of Recursive Titar plan. The main objective is 'a'.

High frequency vital signs + pre-processing before learning forecasting model.

You can see more of those pictures at the gallery page.

Personal area

See the Personal Area page.