Introducing P2C – Partitioning to Classify, an innovative classification technique designed to enhance the performance of linear prediction models on datasets that are not linearly separable. Drawing inspiration from the division-and-conquer strategy, P2C uses clustering methods to partition data into linearly separable groups, enabling the use of linear classifiers with improved accuracy.
Key Features
- Clustering-Based Partitioning: Groups samples of the same class into clusters that are linearly separable.
- Linear Model Utilization: Enables linear classifiers to achieve competitive results on non-linear datasets.
- Scalable: Offers efficient handling of large datasets through parallelization.
- Performance: Comparable or superior to traditional non-linear classification techniques with reduced computational effort.
How It Works
- Clustering by Class: Samples of the same class are clustered to form partitions.
- Union of Clusters: Combines clusters within each partition to create linearly separable groups.
- Training: One or more linear classifiers are trained based on the number of resulting groups.
Publication
- Aumentando o poder preditivo de classificadores lineares através de particionamento por classes
N.A. SOUZA, T.A. ALMEIDA, T.C. SAKATA
Anais do XIV Encontro Nacional de Inteligência Artificial e Computacional (ENIAC’17), Uberlandia, Minas Gerais, Brazil, October, 2017 [pdf]
Code
The source code and documentation are publicly available in this link.