Connectionists: PhD studentship in machine learning in Compiègne (France)

[nicolas.usunier at hds.utc.fr]

  A fully-funded PhD studentship in Machine Learning is available at  
Université Technologique de Compiègne (France).

**Title**

New methods for learning with multiple objectives

**Supervision**
Nicolas Usunier, Associate Professor (nicolas.usunier at hds.utc.fr)
Yves Grandvalet, CNRS Senior Researcher (yves.grandvalet at hds.utc.fr)

**Dates** position open in Fall 2013

**Research Team**

The student will be based in the Heudiasyc laboratory in Compiègne
(France) and join the DI team headed by Yves Grandvalet. He/she will
be supervised by Nicolas Usunier (www.hds.utc.fr/~nusunier/) and Yves
Grandvalet(www.hds.utc.fr/~grandval/). Heudiasyc is a
joint laboratory with the Université de Technologie de Compiègne (UTC)
and the French governmental agency for research (CNRS). In 2011, it
was rated A+ (the highest rate) by the French Research evaluation
agency (AERES). Heudiasyc fosters interdisciplinary research on
information science and technology including machine learning,
uncertain reasoning, operations research, robotics and knowledge
management. In 2011 Heudiasyc was awarded with an excellence project
(LabEx) on the « Control of Technological Systems of Systems ».

**Context**

Most learning algorithms are designed in a risk minimization
framework, where the risk is defined as the expectation of a
task-dependent cost function on the data distribution. While machine
learning algorithm are nowadays applied on a variety of tasks and
large-scale datasets, it turns out that the only algorithms for which
strong theoretical guarantees are proved correspond to the simplest
cost functions for the tasks of classification, regression or
structured prediction.

In real-life applications however, the final performance of a learning
system is often measured by more complex indicators of the true
end-user satisfaction. Examples of such performance indicators include
non-trivial trade-offs between per-class precision and recall (like
macro-averaged F1 scores) in multiclass classification, or trade-offs
between relevance and diversity in search engines or recommender
systems. There is currently no learning algorithms that can provably
and efficiently optimize such performance indicators.

The starting point of the PhD subject is that the principle of
minimizing a single real-valued cost function either (1) does not give
sufficient degrees of freedom to specify all aspects of the
performance of a prediction function, or (2) leads to cost functions
with a complex structure, which cannot be optimized with usual
(e.g. convex optimization) approaches.

**Subject**

The goal of the project is to design new multi-objective approaches to
machine learning, in order to develop methods that can optimize
performance indicators that are non-trivial trade-offs between
different cost functions. The intended results are the design and
analysis of convex multi-objective machine learning algorithms to
problems where convex single-objective approaches are
inadequate. Experimental studies can be conducted in different
application domains, such as search engines, recommender systems or
technological systems of systems in collaboration with other
researchers of the laboratory.

**Requirements**

The PhD candidate should have or expect to obtain a MSc or equivalent  
in computer science or mathematics. The following qualities are  
desirable : strong interests in machine learning or statistics ;  
excellent record of academic and/or professional achievement ; strong  
mathematical skills ; strong programming skills ; good written and  
spoken communication skills in French or English.

**Contact and Application**

Applicants should send a CV and a brief statement of research  
interests by e-mail at nicolas.usunier at hds.utc.fr.