Gradual Shrinkage of Feature Space using ANOVA for WiFi-based Indoor Localization

It is high time researchers realized the potential of statistical inference in setting expectations from features in prioritizing them. In this age of fanatic developments in artificial intelligence, no model would live up to satisfaction in case the features are dismal and insignificant. So, an insight into the significance of features may come to aid prior to delving into arduous hyperparameter tuning. In this paper, we explore the problem of indoor localization and navigation-feeding to an eternal inquisitiveness of humans about where they are at some unknown venue-showing how inferential ANOVA (Analysis of Variance) can shed off noisy features. We utilize ANOVA for double-thresholding, with the first phase eliminating features with p-values > 0.05 and the second round discarding features with F-ratios < 70, across six indoor locations. This scaled down the feature-count from an initial 127 to an intermediate 125 to a final 50-making no compromise in accuracy or F1-score. We cross-validate the shrunk feature-set using kNN, assuming this will render more reliable metrics than neural networks for the purpose. The approach has shown extendibility to multiple classes and has enabled numerical quantification of the impact of each continuous feature-empowering the decision whether to retain the feature, to the alleviation of the curse of dimensionality.