Unsupervised Classification of Categorical Time Series through Innovative Distances

Abstract

- In this paper, two novel distances for nominal time series are introduced. Both of them are based on features describing the serial dependence patterns between each pair of categories. The first dissimilarity employs the so-called association measures, whereas the second computes correlation quantities between indicator processes whose uniqueness is guaranteed from standard stationary conditions. The metrics are used to construct crisp algorithms for clustering categorical series. The approaches are able to group series generated from similar underlying stochastic processes, achieve accurate results with series coming from a broad range of models and are computationally efficient. An extensive simulation study shows that the devised clustering algorithms outperform several alternative procedures proposed in the literature. Specifically, they achieve better results than approaches based on maximum likelihhod estimation, which take advantage of knowing the real underlying procedures. Both innovative dissimilarities could be useful for practitioners in the field of time series clustering.