Comments on real-valued negative selection vs. real-valued positive selection and one-class SVM

Abstract

Real-valued negative selection (RVNS)is an immune-inspired technique for anomaly detection problems. It has been claimed that this technique is a competitive approach, comparable to statistical anomaly detection approaches such as one-class Support Vector Machine. Moreover, it has been claimed that the complementary approach to RVNS, termed real-valued positive selection, is not a realistic solution. We investigate these claims and show that these claims can not be sufficiently supported.

Novelty detection in a changing environment: A negative selection approach

“…The second concern with dimensionality relates to geometry; suppose that one considers the complement of the self set to be a hypercube with each edge dimension spanning the anticipated range of one of the features. It is argued in [16] (convincingly, it must be said) that this is complicated in high feature-space dimensions by the counter-intuitive nature of hyperspheres and hypercubes. However, one of the examples considered in [16] has a 784-dimensional feature-space and the one considered here has a comparatively modest dimension and it is therefore considered viable to apply negative selection.…”

“…It is argued in [16] (convincingly, it must be said) that this is complicated in high feature-space dimensions by the counter-intuitive nature of hyperspheres and hypercubes. However, one of the examples considered in [16] has a 784-dimensional feature-space and the one considered here has a comparatively modest dimension and it is therefore considered viable to apply negative selection. An alternative to negative selection suggested in [15,16] is positive selection; this takes the alternative approach of covering the self set with detectors (hyperspheres) and then flagging novelty if a new, monitored point fails to fall within the detector set.…”

“…As regards the matching threshold for the antibody-set data, each detector x_d i has its own matching threshold r d i calculated following the approach of [13,14]. As discussed in [15,16], the effectiveness of the negative selection algorithm will be related to how well the antibody set D covers the complement of the self set S, as the novelty detector fires if a new monitored point is found to be 'close' in some sense to a member of the antibody set. For simplicity, it will be assumed for now that closeness is measured in terms of Euclidean distance to a member of the antibody set rather than cosine similarity.…”

“…The resulting negative-selection algorithm has already been applied in the literature to the problem of damage detection in [13,14], which are based on time-series features. The suitability of the negative selection algorithm has been discussed in the literature recently in [15,16] and concerns have been raised; largely in relation to the dimensionality of the feature spaces, which can be accommodated by the method. As the dimension of the feature space is modest compared to that discussed in [16]; it is considered that negative selection is worth considering.…”

“…The suitability of the negative selection algorithm has been discussed in the literature recently in [15,16] and concerns have been raised; largely in relation to the dimensionality of the feature spaces, which can be accommodated by the method. As the dimension of the feature space is modest compared to that discussed in [16]; it is considered that negative selection is worth considering. The current paper has a number of modest objectives: the first is to explore the use of the algorithm in the context of general feature vectors for Structural Health Monitoring (SHM), specifically for features based on transmissibilty functions, which have previously proved to be powerful [17][18][19][20].…”