Projects

A Machine Learning Approach for Default Rates and Credit Recovery Intent

Leader: Jean Paul Barddal

Researchers:

Students: Bruno Yuiti Leão Imai, Cristiano Mesquita Garcia, Marcio Vinicius de Souza da Rocha

This project is financed by CNPq and is managed by Softex. In collaboration with Monest, we tackle different aspects regarding credit recovery, including: mapping default rates across different country regions and predicting/scoring the credit recovery intent of individuals. Our ultimate goal is to provide tools to better understand and analyze default and credit recovery across different country regions using different types of data, i.e., socioeconomic, macroeconomic, stock market, and individuals’ data.

Funding and Partners

ChampAnalytics - Predictive Learning Analytics

Leader: Carlos Silla Jr. and Emerson Cabrera Paraiso

Researchers:

Students: Luiz Bernardo Kummer

This research and development project aims to develop a tool for predicting the students’ grades and probability of leaving an undergraduate course.

Funding and Partners

CSCW-SD - An architecture for supporting small collocated teams in cooperative software development

Leader: Emerson Cabrera Paraiso

Researchers: Cesar Augusto Tacla (UTFPR)

Students: Felippe Alex Scheidt

This research project targets studying and developing tools for enhancing collaborative works in software development. In particular, we are interested in the development of novel tools for accompanying and enhancing source code quality.

Funding and Partners

Multimodal/Multiview Learning Applied to Real Problems

Leader: Alceu de Souza Britto Jr.

Researchers: André Gustavo Hochuli, Jose Saavedra, Laurent Heutte, Jean Paul Barddal, George Darmiton da Cunha Cavalcanti, Luiz Eduardo Soares de Oliveira

Students: Carlos Magno Nunes, Victor Schneper Lacerda

This STICAMSUD project consolidates scientific exchange and collaboration of research teams from France, Brazil, and Chile, started in SticAmsud 2014. The main suject is multimodal/multiview machine learning models. We perceive the world around us in a multimodal way. Consequently, multimodal systems have received much attention in Artificial Intelligence (AI). The challenges in such an approach go far beyond the simple fusion of modalities. In practice, it is relevant to consider: a) the modalities to be used, b) the representation of each modality, c) the way modalities interact, and d) how learning accounts for different modalities (individually or simultaneously). This proposal aims to enhance our understanding of the strategies above and develop novel multimodal and multiview predictive models, investigating alternatives for representing and combining different modalities in Pattern Recognition and Machine Learning. The project specifically focuses on researching and developing solutions using multimodal prediction models, emphasizing the architecture of these models, new strategies for generating representations for each modal (views), and new strategies for the selection/fusion of views and modalities considering possible intermodal relationships. Additionally, the project comprises the investigation of strategies to improve the training process of multiple modalities. Each stage of the working plan has the potential to result in publications in qualified international scientific journals and conferences. The project’s social contribution relates to human resource formation and the potential for generating technological innovations that can positively impact society in terms of emotion analysis to help in the diagnosis of depression, medical image analysis, document image retrieval in digital libraries, and car safety (driver assistance systems).

Funding and Partners

Aprendizagem de Máquina Explorando a Interação de Múltiplas Modais

Leader: Alceu de Souza Britto Jr.

Researchers: Paulo Ricardo Lisboa de Almeida, Jean Paul Barddal, George Darmiton da Cunha Cavalcanti, Luiz Eduardo Soares de Oliveira, Robert Sabourin, Rafael Menelau

Students: Luciana Menon, Israel Laurensi, Daniel Vriesman, Luiz E. Soares

Project approved under CNPq’s call for proposals to strengthen cooperation with proven international collaboration, in our case, with researchers from the École de Technologie Supérieure (ÉTS, Canada) that began in 2001. The goal of the research is to develop new multimodal and multi-representation approaches for predictive models based on machine learning techniques aimed at solving classification and regression tasks. To achieve this, three key topics will guide the research to be conducted: (i) the architecture of systems with multiple modalities and multiple representations, (ii) the use of representations based on unsupervised learning, and (iii) the search for suitable strategies to fine-tune deep models for application in new domains (continuous learning). The multimodal applications will be varied, with a preference for ongoing projects with the aforementioned international partnerships, related to emotion recognition (facial expressions and voice), solutions to enhance vehicle safety and monitor the vehicle’s external environment, content-based image retrieval, and medical image analysis.

Funding and Partners

Desenvolvimento de modelo multivariado para previsão de demanda de derivados de petróleo com foco em escalabilidade e produtificação

Leader: Alceu de Souza Britto Jr., Vinicius M. A. Souza

Researchers: Jean Paul Barddal

Students: Alex Cohen Dambros Lopes, Alexandre Beiruth, André Gustavo da Rosa Ribeiro, Jonas Krause, Lucas Gabriel Mendes de Castro

This project is part of the Large-Scale Multivariate Model Development Program for Forecasting Petroleum Derivative Demand, developed in partnership with CISIA (PUCPR Artificial Intelligence Center, a laboratory accredited by the National Petroleum Agency). Its main objective is to create artificial intelligence and machine learning techniques focused on analyzing large sets of time series derived from data provided by the National Petroleum Agency.

Funding and Partners

Machine Learning Aplicado a Previsão de Doenças Graves em Clientes de Saúde Suplementar

Leader: Fabrício Enembreck

Researchers: Júlio Cesar Nievola

Students: Maria Clara Silk

The main objective of this project is the development, evaluation, and application of predictive models for forecasting high-risk cardiovascular diseases such as ischemic heart diseases and strokes using databases maintained by Unimed.

Funding and Partners