23 August 2024
09:00 Doctoral defense fully remotely
Theme
Regression based on Re-ranking
Student
Filipe Marcel Fernandes Gonçalves
Advisor / Teacher
Ricardo da Silva Torres - Co-supervisor: Daniel Carlos Guimarães Pedronette
Brief summary
Several regression-based approaches have been developed in recent years with the aim of improving prediction results, including the use of ranking strategies. Ranking algorithms are mainly based on the calculation of distance functions that compare pairs of points in a high-dimensional feature space. However, pairwise analysis often fails to consider more global similarity relationships. A promising line of research reflects the use of unsupervised re-ranking approaches, whose main objective is to improve the effectiveness of the analyzed rankings. Such methods aim to explore encoded information about the intrinsic structure of the data set. Re-ranking methods redefine the distances between elements in a dataset, leading to more effective rankings, in which relevant objects are allocated in the first positions. Re-ranking methods are widely explored and successfully used in several applications, improving ranks by coding the data structure and redefining distances between elements in a set. Generally implemented as iterative procedures, re-ranking methods redefine the distances between objects, leading to more effective rankings, in which more relevant objects are allocated in the first positions. Despite promising results observed, re-ranking algorithms have not yet been evaluated on regression tasks. This thesis introduces two new contributions with the aim of exploring re-ranking methods for regression, and its main objective is to reduce prediction errors. The first contribution is the proposal of a new, generic and customizable framework entitled Regression by Re-ranking (RbR). This approach exploits the ability of re-ranking algorithms to determine relevant ranks applied to prediction tasks. The RbR framework is built upon the integration of some factors: a base regressor; unsupervised re-ranking learning techniques; and predictions associated with the nearest neighbors weighted according to their position in the ranks. RbR was evaluated under a rigorous experimental protocol including 14 datasets and 32 regression-based methods. The use of RbR achieved significant gains (up to 79%) when compared to state-of-the-art approaches. The second contribution of this work is based on the proposal of a new method for the fusion of regressors, called Fusion of Regressors (Fusion Regression -- FuR), which also explores the neighborhood context. FuR employs ensemble learning that combines the predictions of different regressors. First, different regressors are trained. Then, FuR creates a new embedding based on the regressors' predictions. The objective is to restructure the data set and also explore the complementary insight provided by the regressors. This new representation of the data is then used as input for regression methods that explore the neighborhood context.
Examination Board
Headlines:
| Ricardo da Silva Torres | IC / UNICAMP |
| Henrique Murilo Gaspar | NTNU/Norway |
| Guillermo Cámara Chávez | DECOM/UFOP |
| Guilherme Palermo Coelho | FT / UNICAMP |
| Marco Antonio Garcia de Carvalho | FT / UNICAMP |
Substitutes:
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| João Paulo Papa | FC / UNESP |
| Moacir Antonelli Ponti | ICMC / USP |