A Density-Ratio Approach to Machine Learning
Abstract
A new method of density-ratio estimation that can avoid density estimation was developed. This method gives the solution analytically just by solving a system of linear equations, so it can be applied to large-scale problems. Statistical properties of density ratio estimators have also theoretically investigated. Based on these results, practical machine learning algorithms were developed which include outlier detection, supervised dimensionality reduction, causal direction inference, independent component analysis, conditional density estimation, probabilistic classification, and their performance are shown to be comparable to the state-of-the-art. These algorithms have been applied to solve several real-world problems, which includes speaker identification, audio tagging, nonstationarity adaptation in brain-computer interface, efficient sample reuse in robot control, active exploration in robot control, feature selection in robot control, adaptation of lighting-condition change in face-based age recognition, detection of regions of interest in images , and multi-user adaptation in accelerometer-based human activity recognition.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 20, 2011
- Accession Number
- ADA541193
Entities
People
- Masashi Sugiyama
Organizations
- Tokyo Institute of Technology