FCODE: a Pseudo Data Labeling Scheme for Sensing Systems Using Compressive Sensing
Compressive Sensing (CS) is a novel data representation technique which can represent sparse signals below the Nyquist rate. Since CS can encode multi-dimensional data including images, acceleration, gyro and other sensor data into small representation with a small computation costs, many recent works utilize it to gain their performance. In this work, we utilize CS for multi-dimensional sensing database to both reduce storage usage and improve utility. Multi-dimensional sensing data is likely to occupy the storage and thus, data compression is an effective solution. However, context information cannot be extracted from the compressed data without decompression. This situation causes a huge disadvantage when mining through the sensed data. Motivated by this, we propose FCODE: a Pseudo Data Labeling Scheme for Sensing Systems using Compressive Sensing. FCODE aims to reduce the data dimension while loosely maintaining the meanings of the original data. Although dimension reduced data do not perform as well as the original raw data, reduced data possess adequate ability to apply the first step algorithm of data mining such as classification and feature extraction. We design a hybrid measurement matrix that consists of Gaussian measurement matrix and feature measurement matrix. The former matrix processes for good accuracy in reconstruction and the latter matrix attempts to loosely maintain the meanings of the original raw data. We evaluate our scheme in terms of reconstruction accuracy and classification accuracy
as a performance of feature measurement matrix in simulation and real world based experiments. As a result, the difference of exact accuracy ratio between FCODE and Gaussian matrix is only 0.11% in the best case and FCODE also achieves, in the best case, 88.33% classification accuracy that is close to those using the original signal.
key words: Compressive Sensing, Machine Learning, Pattern Classification, Sensor Database
得られた研究成果は、修士論文としてまとめたほかに、電子情報処理学会英文論文誌の特集号、 "Emerging Technologies on Ambient Sensor Networks toward Future Generation" に投稿した。現在査読中であり、採択された場合は2015年9月に出版される予定である。