Abstract
Feature selection and instance selection are dual operations on a data matrix. Feature selection aims at selecting a subset of relevant and informative features from original feature space, while instance selection at identifying a subset of informative and representative instances. Most of previous studies address these two problems separately, such that irrelevant features (resp. outliers) may mislead the process of instance (resp. feature) selection. In this paper, we address the problem by doing feature and instance selection simultaneously. We propose a novel unified framework, which chooses instances and features simultaneously, such that 1)all the data can be reconstructed from the selected instances and features and 2) the global structure which is characterized by the sparse reconstruction coefficient is preserved. Experimental results on several benchmark data sets demonstrate the effectiveness of our proposed method.
Highlights
Real world applications usually involve with big data with large volume and high dimensionality, presenting great challenges such as the curse of dimensionality, huge computation and storage cost
We propose a novel method for dual selection, which chooses instances to reconstruct all the data in the reduced feature space and keeps the best features to preserve the global structure characterized by the sparse reconstruction coefficients among the selected instances
EFFECT OF DUAL SELECTION STRATEGY Here, we investigate the effect of dual selection within one unified framework by empirically answering the first question as follows
Summary
Real world applications usually involve with big data with large volume and high dimensionality, presenting great challenges such as the curse of dimensionality, huge computation and storage cost. A number of methods [1], [14]–[22] aim to exploit the intrinsic cluster structure of data, and use it as pseudo label for further feature selection task Another line of work [15], [23]–[31] is to select those features which can be used to well reconstruct or approximate the whole data set. Since the above problem is NP-hard, UFI uses a greedy algorithm to solve it, where the importance of each feature and instance is evaluated individually and less informative feature (instance) is removed one by one Such backward removal mechanism does not take special consideration on the correlations between features and instances. The difference of our method and UFI are as follows, 1) different formulations; 2) UFI evaluates the importance of each feature and instance one by one, while our method can evaluate a subset of features and instances jointly, which leads to better performance
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
