Weighting features by the value displacement rebound

Abstract

Learning from examples draws on similarity, a concept which formalisation leads to the notion of instance space. Continuous spaces are easier to embrace since, unlike discrete, they often can be seen as hyper-constructs of 3D. Unsurprisingly, the instance-based learning methods are more developed for continuous domains than for discrete ones. The value difference metric (VDM) is one of the few examples of metrics for discrete spaces. Mixed reports about utility of VDM exist. In this paper VDM is compared with another approach where data features are weighted by the Information Gain. Some vulnerabilities of VDM are identified. A weighting method, nothing like VDM, although inspired by the former, is proposed. The results are in favour of the new weighting scheme with illustration of utility for health diagnostics.

Decision branch joint venture ex-fold T-z re-validation

“…The proposed lazy DT is better equipped than the standard one to deal with a possible class imbalance by absorbing the contained in leaves secondary information passed on to them by progenitor nodes. However, the Table 2 results for DB-JV, while acceptable, do not compare favorably with previously reported ones by other methods for the same data, [18] particularly the earlier mentioned Info-Neighbor which uses IG weighted overlap metric from Eq.13.…”

“…This is the same data-set as previously reported. [18] The surveys are conducted two-yearly. Each survey is an independent snap-shot of the population.…”

“…Evidently, J48 lacks capabilities to deal with the class imbalance. [18] So, at least for DM, the balanced accuracy results in Table 2 are much different.…”

“…[15] The results in Tables 1&2 are consistent with the theoretical accuracy variance being smaller for larger samples, according to Eq.1. It is not possible to do calculations exactly without adapting Eq.1 for the balanced accuracy, but this metric approximates the situation when samples are classbalanced, [18] that is, there are equal numbers of any class. By seeing Eq.2 in this light it is easy to verify that Eq.1 can then be used as is.…”

“…For example, with NB, modifying the training set by removing one instance and adding another will not cause a substantive difference to the involved feature value probabilities by class in a large data. [18] Some apprehension about LOOCV exists, though, and this may be because of its proximity to resubstitution. [2,3] Theoretically, LOOCV may fail on noisy data, due to overfitting, if the classifier is unstable.…”