KNMFS: a novel unsupervised feature selection method
Date:
This was my guest lecture for the Roda de Conversa de Dados at CESAR. I introduced the audience to the concepts and challenges of feature engineering and presented KNMFS, a novel unsupervised feature selection method derived from my research.
Topics covered:
- The definition of features and their inherent challenges, including data dispersion and the curse of dimensionality.
- Continuous Integration in Data Science workflows, focusing on Continuous Delivery for Machine Learning (CD4ML) and Feature Stores.
- An in-depth overview of the KNMFS (Kernelized NMF and Modularity Feature Selection) method, which is based on iterative similarity graph factorization and clustering by modularity.
- The five-step KNMFS pipeline:
- 1. Similarity Matrix construction.
- 2. Iterative Non-negative Matrix Factorization (NMF) with a Gaussian filter.
- 3. K-Nearest Neighbors (KNN) Similarity Graph generation.
- 4. Louvain Community Detection for modularity optimization.
- 5. Feature Scoring using L2-norm.
- Experimental results on real and synthetic datasets demonstrating the performance of KNMFS against state-of-the-art methods regarding Accuracy, NMI, and ARI metrics.