Unsupervised group feature selection for media classification

Abstract

The selection of an appropriate feature set is crucial for the efficient analysis of any media collection. In general, feature selection strongly depends on the data and commonly requires expert knowledge and previous experiments in related application scenarios. Current unsupervised feature selection methods usually ignore existing relationships among components of multi-dimensional features (group features) and operate on single feature components. In most applications, features carry little semantics. Thus, it is less relevant if a feature set consists of complete features or a selection of single feature components. However, in some domains, such as content-based audio retrieval, features are designed in a way that they, as a whole, have considerable semantic meaning. The disruption of a group feature in such application scenarios impedes the interpretability of the results. In this paper, we propose an unsupervised group feature selection algorithm based on canonical correlation analysis (CCA). Experiments with different audio and video classification scenarios demonstrate the outstanding performance of the proposed approach and its robustness across different datasets.

A systematic review and evaluation of statistical methods for group variable selection

“…One such situation is the analysis of variables that are interrelated based on correlations or contextual similarities, since ignoring such group structure reduces the stability, consistency, and interpretability of the selection 3 . Methods accounting for group structures have been investigated since at least 1999 4 and have been used in a broad range of applications, including media classification, 5 disease prediction, 6 automotive engineering, 7 voting behavior analysis, 8 emotion recognition, 9 and credit risk analysis 10 . One of the most common applications is in omics research, such as gene expression microarray or single nucleotide polymorphism data 11‐15 .…”

Group Based Unsupervised Feature Selection

“…Many unsupervised feature selection methods, both similarity preserving (filter) [9,19] and embedded [6,8,10,14] methods, are based on input data alone and rarely take the advantage of the external sources of knowledge about feature group structures. The feature groups used by some feature selection methods are also formed with input data [15,18]. Some domain specific unsupervised methods [3] are proposed for selecting genes from different gene groups, yet they do not combine group based feature selection with instance-feature data which is also useful for feature selection.…”

Interpretability of bi‐level variable selection methods

“… Collinearity tolerance : Highly correlated predictors are treated alike since it is assumed that high correlation implies similarity in content, and interpretability is enhanced if those variables are selected jointly (Zou & Hastie, 2005; Bondell & Reich, 2008; Dormann et al., 2013). Group‐level consistency : Once a feature of a prespecified group is added to the model, all variables of that group are included, since it is likely that the predictors of a group are only meaningful together (Zaharieva et al., 2017; Yuan & Lin, 2006; Breheny & Huang, 2015; Gregorutti et al., 2015). …”

“…When predictors form groups based on their correlation or contextual similarity, related information is spread over numerous features. This makes selection at the group level more appropriate than identification of individual variables (Subrahmanya & Shin, 2010; Zaharieva et al., 2017). Example scenarios are when a set of features was collected with the same measurement instrument, was derived in different ways from the same data, or originated from the same experimental setting.…”

“…Group‐level consistency : Once a feature of a prespecified group is added to the model, all variables of that group are included, since it is likely that the predictors of a group are only meaningful together (Zaharieva et al., 2017; Yuan & Lin, 2006; Breheny & Huang, 2015; Gregorutti et al., 2015).…”