When: May 21, 2013 from 3PM to 4:30PM

Where: room B11

Speaker: Jesse Read

Title: Advances in Multi-label Classification

Abstract: In multi-label classification, each data instance may be associated with multiple classes, as opposed to a single class label. The multi-label context arises naturally in many domains, such as text categorisation and labelling images. The main challenge is modelling dependencies between labels, which must be done efficiently to scale up to settings involving large datasets and data streams. This talk reviews recent advances in multi-label classification specific to the approach of ‘classifier chains’, and also discusses methods for multi-label learning in the context of data streams.

When: April 11th, 2013  from 11AM to noon.

Where: room B21

Speaker: Cristina Manfredotti, LIP6 UPMC (Paris 6)

Title: Relational Inference and Learning for Complex Recognition Tasks

Abstract:
Many domains in the real world are richly structured, containing a multitude of entities related to each other in a variety of ways. Many problem domains require modeling the behaviors of multiple agents, understanding their roles, the context and detecting anomalies. Examples of such domains span from surveillance systems (in which, for example, one has to identify the activity of multiple interacting agents) to systems for marketing support (discount campaigns targeting specific key customers, product recommendations based on similar purchases made by others, etc.) and from bio-informatics (relationships between patients’ genetic profiles and their drug responses) to human motion understanding (relationships between active components, such as joints in body motion analysis). A key characteristic of many situations is that interactions (or relations) are dynamic and may change over time. The explicit recognition of the relationships between entities can improve the understanding of their behaviors, and help predicting future trends.

In this talk we will cover the problems of (1) modeling dynamically changing relations between entities, (2) making inference in dynamic domains taking into account relations between entities and (3) learning probabilistic models of interaction of entities. In order to automatically learn probabilistic models that take into account dinamically changing relations we propose a semi-supervised learning framework that makes use of hierarchical abstraction. Experimental results on the particular application of multi target tracking and (online) activity recognition will be considered. These show that our approach decreases the tracking error rate, improves the data association performance and identifies the correct activity with higher accuracy than standard approaches.

When: April 4th, 2013  from 11AM to noon.

Where: room B21

Speaker: Antoine Bordes (CR CNRS) from Heudiasyc (UT Compiègne)

Title: New models for highly multi-relational data

Abstract:
Many data such as movie preferences or knowledge bases are
multi-relational, in that they describe multiple relations between
entities. While there is a large body of work focused on modeling these
data, modeling these multiple types of relations jointly remains
challenging. Further, existing approaches tend to breakdown when the
number of these types grows. In this talk, we present methods for
modeling large multi-relational datasets, with possibly thousands of
relations. Our models are based on bilinear structures, which captures
various orders of interaction of the data. We illustrate the performance
of these approaches on standard tensor-factorization datasets where we
attain, or outperform, state-of-the-art results. Finally, a NLP
application demonstrates our scalability and the ability of our models
to learn efficient and semantically meaningful word representations.

When: Friday (March 15th, 2013) from 11AM to noon.

Where: room B21

Speaker: Emmanuel Viennet (PU université Paris 13)

Title: Analyse des réseaux sociaux et des communautés — Exemples d’applications

Abstract:
L’analyse des réseaux sociaux et des graphes complexes font l’objet d’un grand intérêt ces dernières années, poussé par les enjeux industriels et les défis posés par le traitement des données massives, liées, hétérogènes, bruitées générées par les applications. Lors de cet exposé, nous présenterons quelques méthodes développées
ces dernières années pour la recherche de communautés dans les graphes sociaux: communautés locales ou ego-centrées, prise en compte simultanée des liens et des attributs.
Ces méthodes seront illustrées par quelques applications récentes issues de collaborations industrielles.