Research

Research Topics
Vodkaster/ALICIA
Project Imagiweb
Evaluation Campaigns
Ph.D. thesis
Master thesis

Research Topics

• Natural Language Processing
• Machine Learning
• Information Retrieval

At Vodkaster I was mainly working on movies recommender systems within the ALICIA (ANR-13-CORD-0020) project. My work mainly consisted in studying features to improves recommendation. It also included modelling and analysis dynamics of movies reputation on the website. My research area also included Network broadcasting and Video coding. I am also interested in modellig and improving customers' satisfaction within network analysis.

Subject: Analysing entities web representation over Web 2.0
PhD Supervisor: Marc El-Bèze, Juan-Manuel Torres-Moreno and Eric Sanjuan
My main work consists in is about entities reputation representation identification. There are two major intentions: analyse and visualisation.
The web and politics context add difficulties to the analyse side. On the web there many bias, the first one could be that the web is offering more coverage to negative messages. Politics is like shifting and people can easily change their opinion and each new fact become an emerging concept. When it comes to modeling of the reputation the contexts add other difficulties such as the point-of-view and how to estimate the right (temporal) window. The project includes different working parts, from manual annotation with domain experts to machine learning with automatic annotation and clustering.

This project is highly related to the domain of Sentiment Analysis, and more specifically to Opinion Mining. The main idea is to detect “what people think about a given entity” from documents content. This idea is usually implemented within a more realistic task: classifying the opinion expressed about something into a set of predefined polarities (e.g., positive vs. negative or neutral). The nature of vocabulary used in tweets limits the use of existing Sentiment Lexicon. This project also aims to study on which aspect of the entity the opinion was expressed. Although it could be seen as Topic Detection it is harder since the topic set has been defined by experts as a concept level which may never be seen in the contents.
Politics has already been addressed in previous works but mostly in English that deals with the US politics and rarely with the precision expected in the project. Furthermore, a dataset built with the involvement of specialists in political science will be provided to the community.

Within the Imagiweb project my work consists in modelling and analysis dynamics of web reputation. Our objectives aim to help politics researchers and communication-team of the French main electric utility company (EDF) by proposing them tools that can provide automatic topic and opinion annotation on documents dealing with their reputation (tweets about politicians and about the “EDF” entity). The project covers several issue such as Active Learning, Natural Language Processing in Big Data Mining. My research activities cover document clustering and categorization using several Machine Learning methods. I also studies the impact of external features in the retweet behaviour.

RepLab'2014
RepLab 2014 provides a extension to RepLab 2013 tasks. It focused on stress categorization task (using Reputation Standards from the RepTrack Framework) and the characterization of Twitter profiles (Author Profiling) as a complement to CLEF PAN challenge. We tried to tackle both problems with statistical NLP-based classifiers with the simplicity and re-usability philosophy considering a matching dogma as mainstream process.
We obtained competitive results in Author Profiling subtasks.
The Reputation Dimensions classification task looks like the Topic Categorization aspect of the Imagiweb project that's why we made further experiments
My work consisted in team management and participation to all subtasks.

RepLab'2013
RepLab 2013 is an evaluation challenge focusing on the problem of monitoring the reputation of entities in Twitter. It consists in several tasks such as entity name disambiguation (Is the tweet about the entity?), reputation's polarity detection (Does the tweet have positive or negative implications for the entity’s reputation?), topic detection (What is the issue relative to the entity is discussed in the tweet?) and topic ranking (Is the topic a reputation alert that deserves immediate attention?). The provided dataset contained tweets in two languages: English and Spanish. We mainly tried to investigate how much Speech Recognition and Information Retrieval systems can answer the issues in a reputation management context (filtering and polarity tasks 1 and 2) and how simple NLP-based classifiers can perform over ranking and clustering tasks (task 3 and 4). We obtained competitive results in each subtasks.
My work consisted in team management, merging all systems and of course participation to several subtasks with my own ideas.

Deft’2013
Deft 2013 edition addressed a new application domain on a theme that has been studied in an evaluation campaign in the past (Computer Cooking Contest): cooking recipes. We focused on two analysis functions in DEFT2013, document classification (task 1 to 3) and information extraction (task 4), in a speciality domain.
My participation consisted to be a expert annotator to evaluate subsets of systems submission as if we were in an active learning process.

Subject: Opinion mining in a movie recommender system

Master Supervisor: Marc El-Bèze

This Master thesis studies an opinion mining system over a movies social network. This method relies on Natural Language Processing over the users reviews. Several aspects were covered, from the user point-of-view: what does he like or dislike in movies and from the movie side: what the main target of the movie. This analyse is used to propose an argued movie recommendation for each user.

Publications

International Conferences

Journal
International Workshop

National Conferences

Challenges
National Workshop
Others

Posters

Slides

Teaching

Year 2018-2020 Teaching

Courses & Practical Work	Hours
Data Minig (M.Sc.)	2x 12
Digital Shadow (M.Sc.)	2x 12
Data Governance (M.Sc.)	2x 12
Total	72

194 teaching hours during PhD, including Practical work &Tutorial (Detailed information):

Year 2014/2015 Teaching

Tutorials & Practical Work	Hours
C++ Base Programming (B.Sc.)	17
Advanced C++ Programming (B.Sc.)	25.5
NLP Database integration and graph analysis in Social Networks (M.Sc)	21
Internship Supervisor (M.Sc)	-
Project Supervisor (M.Sc) : Hypervisor vs Docker	-
Total	63.5

Year 2013/2014 Teaching

Tutorials & Practical Work	Hours
C++ Base Programming (B.Sc.)	24
Advanced C++ Programming (B.Sc.)	25.5
Network infrastructure & Social Networks (M.Sc)	21
Internship Supervisor (M.Sc) : PLS-PM NLP in football betting	-
Total	70.5

Year 2012/2013 Teaching

Tutorials & Practical Work	Hours
C++ Base Programming (B.Sc.)	20
Advanced C++ Programming (B.Sc.)	24
Network Programming (M.Sc)	21
Project Supervisor (M.Sc) : Deep Learning NLP	-
Total	65

CV
(Researcher / Data Officer-Analyst-Scientist / Research Engineer)

Since 2017 January, I work at My Local Influence as Research Engineer. My main concerns deal with Online Reputation Management. After about 1 year, I left my Research Engineer position at Vodkaster where I mainly worked on recommender systems within the ANR (French National Research Agency) founded project ALICIA.
I used to collaborate at the University of Avignon within the context of the ANR project : Imagiweb about entities' (individuals and companies) reputation analysis over Web 2.0

Resume

Thesis

Research Keyworks

Natural Language Processing
Information Retrieval
Online Reputation Management and Monitoring
Machine Learning
Social Media Analysis
Contents Ranking and Selection (Summarization)
User-generated contents Mining and Categorization
User Profiling (Influence, SCC, Age, Gender, Personality, Political Orientation)
Item Modeling from reviews
Recomender system (or other cultural products)
Artificial Intelligence

Education

09-2012 -- 08-2015 Ph.D. in Computer Science Specialized in Natural Language Processing applied to Online Reputation Analysis, LIA - University of Avignon (France).

09-2010 -- 08-2012 Master of Science, Specialized Networking and Natural Language Processing
CERI - University of Avignon (France).

Employment Experience

10-2015 -- 10-2016 Research Engineer at Vodkaster - Paris (France).

09-2012 -- 08-2015 Lecturer at CERI - University of Avignon (France).
Various courses in Computer Science: Networking, C/C++ and Java (Networking) programming
Introduction to Social Network Analysis.

11-2011 -- 08-2012 Junior Research Assistant at LIA - University of Avignon (France).

Languages

French (Native), English, notions of Italian and Spanish.

Teaching

C++ Base Programming
C/C++ and Java (Networking)
Network and Telecommunications
Social Network Analysis

References

Philippe Fillinger
i-Roe Consulting
Phone: +33 6 16 55 68 15
Email: [email protected]

Chris Navas
Vodkaster / RIPLAY SAS
23 Rue Boyer 75020 Paris, France
Phone: +33 6 20 54 43 14
Email: [email protected]

 

Professor Marc El-Bèze
University of Avignon
339 chemin des Meinajariès
84911 Avignon, France
Phone: +33 490 843 508
Email:[email protected]

Juan-Manuel Torres-Moreno
Associate Professor (HDR)
University of Avignon
339 chemin des Meinajariès
84911 Avignon, France
Phone: +33 490 843 568
Email:[email protected]

Eric SanJuan
Associate Professor
University of Avignon
339 chemin des Meinajariès
84911 Avignon, France
Phone: +33 490 843 568
Email:[email protected]

Contact

45 Rue Frédéric Joliot Curie MYLI
13013 Marseille
FRANCE

[email protected]

+33 665 630 728

Twitter @jvcossu
Google Scholar
DBLP
LinkedIn